正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

XcodeBuildMCP 2.6 is here, and the UI automation tools got a serious glow-up. 🚀 Same task, same simulator, side by side: ⏱ 70% faster (6m57s → 2m06s) 🧠 68% fewer tokens (232.8k → 75k) 🔧 76% fewer tool calls (86 → 21) Watch them race 👇

camsoft2000

3,854 subscribers

28,528 次观看 • 2 个月前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

Still not convinced of the benefits of UI Automation with XcodeBuildMCP? How about having your very own QA tester? The prompt "Please go ahead and test different calculations":

Still not convinced of the benefits of UI Automation with XcodeBuildMCP? How about having your very own QA tester? The prompt "Please go ahead and test different calculations":

camsoft2000

21,670 次观看 • 1 年前

OpenAI's chairman just responded to the Kimi K3 panic 👇 The timeline calls K3 a frontier killer. Bret Taylor's answer on CNBC: every CEO he talks to asks the same 2 questions. 1️⃣ Am I getting value from my token spend? 2️⃣ What's my moat in the age of AI? His counter on cheap Chinese models: frontier models finish the same task in far fewer tokens. Cheap per token means little if you burn 5x more of them The new battleground is tokens per task, and the price war just forced OpenAI to argue efficiency instead of capability

OpenAI's chairman just responded to the Kimi K3 panic 👇 The timeline calls K3 a frontier killer. Bret Taylor's answer on CNBC: every CEO he talks to asks the same 2 questions. 1️⃣ Am I getting value from my token spend? 2️⃣ What's my moat in the age of AI? His counter on cheap Chinese models: frontier models finish the same task in far fewer tokens. Cheap per token means little if you burn 5x more of them The new battleground is tokens per task, and the price war just forced OpenAI to argue efficiency instead of capability

Ruben

76,011 次观看 • 20 天前

I got tired of explaining myself to AI every single day. Same context. Same workflow. Same instructions. Copy-pasted into every new session like it’s 2022. So I tried Panofy. Uploaded my own scripts and documents to train an agent — then gave it a single instruction. 🔥 It ran the task. Got it right. First time. No prompt loops. No setup. It just… knew. The thing that surprised me most? It actually improves the more you use it. Learns from mistakes. Builds memory. Adapts to how you think. Fewer tokens, better results. I’ve used a lot of AI tools. This one made me feel like I finally own mine. Panofy isn’t just another tool — it’s the best AI agent for expert-level tasks. If you’re serious about building with AI, this is worth trying. Sign up and get 10,000 free credits + 7-day Panofy Plus free trial 👉 #AI #AIAgents #Productivity

I got tired of explaining myself to AI every single day. Same context. Same workflow. Same instructions. Copy-pasted into every new session like it’s 2022. So I tried Panofy. Uploaded my own scripts and documents to train an agent — then gave it a single instruction. 🔥 It ran the task. Got it right. First time. No prompt loops. No setup. It just… knew. The thing that surprised me most? It actually improves the more you use it. Learns from mistakes. Builds memory. Adapts to how you think. Fewer tokens, better results. I’ve used a lot of AI tools. This one made me feel like I finally own mine. Panofy isn’t just another tool — it’s the best AI agent for expert-level tasks. If you’re serious about building with AI, this is worth trying. Sign up and get 10,000 free credits + 7-day Panofy Plus free trial 👉 #AI #AIAgents #Productivity

Aryan Rakib

33,551 次观看 • 3 个月前

XcodeBuildMCP running fully natively in Xcode!

XcodeBuildMCP running fully natively in Xcode!

camsoft2000

20,414 次观看 • 6 个月前

I cant believe this guy just made a permanent solution to context bloat and open sourced it all! when we tested this tool (Context+) for solving an issue on the OpenCode repository, the agent using this tool used ~6.5k fewer tokens, found the code and fixed it in half the time! the results were surprising: 6 to 10k tokens saved per prompt, completed task in ~2 minutes while the agent running without the tool took ~4 mins for the same and got stuck in loops bro built an entire beast by using all the modern tools that we could think of: undo trees, semantic search by meaning (by haskellforall), advanced refactoring, blast radius, advanced file context trees, restore points... i can keep going on semantic code search and context trees are the future of agentic coding and this tool proves it the feature i loved the most is semantic search and how it gets things done 2x faster with least possible tokens it makes an agent that actually knows what it’s doing and not just guessing, it makes meaning from your code similar to RAG. if you aren't optimizing your context, you are just burning money the developer says this tool is still under development, it can have unexpected behavior and the docs need updates but the video shows the reality of how fast it can be github: get here:

I cant believe this guy just made a permanent solution to context bloat and open sourced it all! when we tested this tool (Context+) for solving an issue on the OpenCode repository, the agent using this tool used ~6.5k fewer tokens, found the code and fixed it in half the time! the results were surprising: 6 to 10k tokens saved per prompt, completed task in ~2 minutes while the agent running without the tool took ~4 mins for the same and got stuck in loops bro built an entire beast by using all the modern tools that we could think of: undo trees, semantic search by meaning (by haskellforall), advanced refactoring, blast radius, advanced file context trees, restore points... i can keep going on semantic code search and context trees are the future of agentic coding and this tool proves it the feature i loved the most is semantic search and how it gets things done 2x faster with least possible tokens it makes an agent that actually knows what it’s doing and not just guessing, it makes meaning from your code similar to RAG. if you aren't optimizing your context, you are just burning money the developer says this tool is still under development, it can have unexpected behavior and the docs need updates but the video shows the reality of how fast it can be github: get here:

forloop

226,054 次观看 • 5 个月前

This guy lost his private keys to an AI agent. Then his passwords leaked online. He wasn't the only one. Reddit threads blew up. Dozens of people saying the same thing. "My credentials got exposed." "The agent ran code I didn't approve." "I don't trust this anymore." That's the OpenClaw problem nobody talks about. Now there's a fix. It's called IronClaw. Here's what makes it different: Built in Rust → fewer bugs, fewer holes for hackers to exploit Web assembly sandboxes → every tool runs in its own sealed room Encrypted credential vault → the AI never sees your passwords. Ever. Policy enforcement → you set the rules, the agent follows them Full audit logs → see exactly what it did, when, and why Side by side: OpenClaw → credentials leaked, no audit trail, tools accessed everything IronClaw → encrypted vault, sandboxed tools, policy guardrails, logs everything It's not for casual use yet. But if you're running a real business with real data? There's only one right answer here. Go to ironclaw.te and read how the credential vault works. The future of AI agents is secure AI agents. Full stop.

This guy lost his private keys to an AI agent. Then his passwords leaked online. He wasn't the only one. Reddit threads blew up. Dozens of people saying the same thing. "My credentials got exposed." "The agent ran code I didn't approve." "I don't trust this anymore." That's the OpenClaw problem nobody talks about. Now there's a fix. It's called IronClaw. Here's what makes it different: Built in Rust → fewer bugs, fewer holes for hackers to exploit Web assembly sandboxes → every tool runs in its own sealed room Encrypted credential vault → the AI never sees your passwords. Ever. Policy enforcement → you set the rules, the agent follows them Full audit logs → see exactly what it did, when, and why Side by side: OpenClaw → credentials leaked, no audit trail, tools accessed everything IronClaw → encrypted vault, sandboxed tools, policy guardrails, logs everything It's not for casual use yet. But if you're running a real business with real data? There's only one right answer here. Go to ironclaw.te and read how the credential vault works. The future of AI agents is secure AI agents. Full stop.

Julian Goldie SEO

15,686 次观看 • 5 个月前

Coming soon: XcodeBuildMCP debugging a run-time crash with the new LLDB capability. Coding agent can inspect an application at run-time, add/remove breakpoints, get stacktraces, and issue lldb commands just like a human developer would.

Coming soon: XcodeBuildMCP debugging a run-time crash with the new LLDB capability. Coding agent can inspect an application at run-time, add/remove breakpoints, get stacktraces, and issue lldb commands just like a human developer would.

camsoft2000

149,561 次观看 • 7 个月前

your agent has thirty tools. it calls two of them. the other twenty eight are not sitting idle somewhere. they are in the request, every request, and they are doing damage in two places at once. first the obvious one. tool schemas go into the prompt, and a schema is not a name. it is a description, a parameter list, types, required fields, an example. thirty of those is a few thousand tokens that ship with every single call, including the ones where the agent just says thanks and stops. you are paying rent on twenty eight tools that have never fired. second, and this is the one that costs more. when the request says cancel the order, the model picks by matching against everything available. four of your tools are plausible: cancel_order, refund_order, update_order, void_order. it is choosing among them based on the descriptions you wrote, one afternoon, months ago. every tool you add is another candidate in that shortlist. the twenty eight you never call are not neutral. they are noise in the one decision that determines whether the run works. > why it grows without anyone deciding to nobody adds thirty tools on purpose. you add one for a task, it works, it stays. six months later the registry is a catalogue and no one has ever removed anything, because removing a tool feels risky and adding one feels free. and there is no feedback telling you otherwise. the unused ones never error. they never appear in a failing trace. they are invisible in exactly the way that lets them accumulate. > what to actually do count calls per tool over the last thousand runs. this is one group-by and it usually shocks people. the ones at zero are pure cost. ship the tools the task needs, not the whole registry. a research phase does not need deploy. a writing phase does not need the database. swap the set between phases instead of loading everything up front. same agent, different tools, depending on where the run is. and when two tools could both plausibly answer the same request, that is not redundancy you can ignore. it is a coin flip you built into the system. the twenty eight tools are not unused. they are used every time, by the part of the run you cannot see.

your agent has thirty tools. it calls two of them. the other twenty eight are not sitting idle somewhere. they are in the request, every request, and they are doing damage in two places at once. first the obvious one. tool schemas go into the prompt, and a schema is not a name. it is a description, a parameter list, types, required fields, an example. thirty of those is a few thousand tokens that ship with every single call, including the ones where the agent just says thanks and stops. you are paying rent on twenty eight tools that have never fired. second, and this is the one that costs more. when the request says cancel the order, the model picks by matching against everything available. four of your tools are plausible: cancel_order, refund_order, update_order, void_order. it is choosing among them based on the descriptions you wrote, one afternoon, months ago. every tool you add is another candidate in that shortlist. the twenty eight you never call are not neutral. they are noise in the one decision that determines whether the run works. > why it grows without anyone deciding to nobody adds thirty tools on purpose. you add one for a task, it works, it stays. six months later the registry is a catalogue and no one has ever removed anything, because removing a tool feels risky and adding one feels free. and there is no feedback telling you otherwise. the unused ones never error. they never appear in a failing trace. they are invisible in exactly the way that lets them accumulate. > what to actually do count calls per tool over the last thousand runs. this is one group-by and it usually shocks people. the ones at zero are pure cost. ship the tools the task needs, not the whole registry. a research phase does not need deploy. a writing phase does not need the database. swap the set between phases instead of loading everything up front. same agent, different tools, depending on where the run is. and when two tools could both plausibly answer the same request, that is not redundancy you can ignore. it is a coin flip you built into the system. the twenty eight tools are not unused. they are used every time, by the part of the run you cannot see.

Hanako

24,656 次观看 • 2 天前

Tested DeepSeek V4 Flash on ZenMux today Tried a more practical approach instead of just chatting. → Asked it to build a SaaS landing page → Generated full HTML, CSS, JS → Ran the code → got a working UI It handled layout, components, and styling pretty smoothly. Iteration was simple too — just refine the prompt and improve the output. Also liked how easy it is to switch between models in the same place and test things side by side without changing setups. What stood out is how quickly you can go from idea → working frontend without jumping between tools. From a cost perspective, this kind of output is more than enough for a lot of routine dev tasks. Feels less like a “Claude killer” More like a “Claude bill killer.” If you want to try it yourself, I’ve added the link below 👇 Try it here: ZenMux

Tested DeepSeek V4 Flash on ZenMux today Tried a more practical approach instead of just chatting. → Asked it to build a SaaS landing page → Generated full HTML, CSS, JS → Ran the code → got a working UI It handled layout, components, and styling pretty smoothly. Iteration was simple too — just refine the prompt and improve the output. Also liked how easy it is to switch between models in the same place and test things side by side without changing setups. What stood out is how quickly you can go from idea → working frontend without jumping between tools. From a cost perspective, this kind of output is more than enough for a lot of routine dev tasks. Feels less like a “Claude killer” More like a “Claude bill killer.” If you want to try it yourself, I’ve added the link below 👇 Try it here: ZenMux

Shruti Codes

69,532 次观看 • 3 个月前

Anthropic shouldn't have made this free a company doing $47,000,000,000 a year wrote down exactly how they run their AI agents, published the numbers, and charged nobody it's called Graph Engineering: one lead Claude plans a job and hires a swarm of smaller ones, each working its own slice at the same time turns out how many it hires decides everything: → a simple lookup: 1 agent, 3 to 10 tool calls, no swarm → a straight comparison: 2 to 4 workers, 10 to 15 calls each → open-ended research: 10+ workers, one slice of the question each → the lead fires 3 to 5 at once, each running 3+ tools in parallel: up to 90% faster the briefs are where it dies. they told a lead agent to "research the semiconductor shortage" and one worker went off into the 2021 car chip crisis while two others wrote the same 2025 report twice so a worker now gets four things: an objective, an output format, which tools to touch, where its job ends they also pointed one small agent at their own badly written tool descriptions and let it rewrite them every agent that used the new ones finished 40% faster then one grader Claude scores every run 0.0 to 1.0 on five things: factual accuracy, citation accuracy, completeness, source quality, tool efficiency twenty test questions took one of their agents from 30% success to 80% one limit nobody quotes: most coding work has fewer genuinely parallel pieces than research, so a swarm on one repo mostly buys you coordination overhead it pays on wide search and on jobs bigger than one context window free, out of a $965,000,000,000 lab, and almost nobody has copied it yet bookmark this and copy the counts ↓

Anthropic shouldn't have made this free a company doing $47,000,000,000 a year wrote down exactly how they run their AI agents, published the numbers, and charged nobody it's called Graph Engineering: one lead Claude plans a job and hires a swarm of smaller ones, each working its own slice at the same time turns out how many it hires decides everything: → a simple lookup: 1 agent, 3 to 10 tool calls, no swarm → a straight comparison: 2 to 4 workers, 10 to 15 calls each → open-ended research: 10+ workers, one slice of the question each → the lead fires 3 to 5 at once, each running 3+ tools in parallel: up to 90% faster the briefs are where it dies. they told a lead agent to "research the semiconductor shortage" and one worker went off into the 2021 car chip crisis while two others wrote the same 2025 report twice so a worker now gets four things: an objective, an output format, which tools to touch, where its job ends they also pointed one small agent at their own badly written tool descriptions and let it rewrite them every agent that used the new ones finished 40% faster then one grader Claude scores every run 0.0 to 1.0 on five things: factual accuracy, citation accuracy, completeness, source quality, tool efficiency twenty test questions took one of their agents from 30% success to 80% one limit nobody quotes: most coding work has fewer genuinely parallel pieces than research, so a swarm on one repo mostly buys you coordination overhead it pays on wide search and on jobs bigger than one context window free, out of a $965,000,000,000 lab, and almost nobody has copied it yet bookmark this and copy the counts ↓

Argona

116,796 次观看 • 15 天前

Gavin Baker just spotted a massive contradiction inside the AI market. The valuations across the AI supply chain don't add up when you hold them side by side. - Memory names: trading at 3-5x PE. - Nvidia: trading at a very low PE. - Some accelerator companies: reasonable multiples. - Power and cooling names: discounting a much higher future. These multiples can't all be right at the same time. If power and cooling names are priced correctly, then Nvidia and memory are way too cheap. They're going up a lot from here. If Nvidia and memory are priced correctly, then power and cooling will significantly underperform from here. One side is wrong. Both can't be true. Baker's word for it: the AI market is "cross-sectionally inefficient" right now. For investors, the question isn't just "am I bullish on AI?" The real question is: which part of the stack is priced correctly and which part is priced for a scenario that isn't going to play out? Note: Baker runs Atreides Management, one of the most respected tech-focused growth funds in the world. He doesn't make calls like this lightly. When the best investors in tech spot a cross-sectional mismatch this size, the trade is figuring out which side wins. Our analysts are already positioning inside the AI supply chain. They called $AMD, $MU, $CRDO and $NBIS before the big runs. Follow their exact portfolios for $1 at Milk Road PRO, link in bio.

Gavin Baker just spotted a massive contradiction inside the AI market. The valuations across the AI supply chain don't add up when you hold them side by side. - Memory names: trading at 3-5x PE. - Nvidia: trading at a very low PE. - Some accelerator companies: reasonable multiples. - Power and cooling names: discounting a much higher future. These multiples can't all be right at the same time. If power and cooling names are priced correctly, then Nvidia and memory are way too cheap. They're going up a lot from here. If Nvidia and memory are priced correctly, then power and cooling will significantly underperform from here. One side is wrong. Both can't be true. Baker's word for it: the AI market is "cross-sectionally inefficient" right now. For investors, the question isn't just "am I bullish on AI?" The real question is: which part of the stack is priced correctly and which part is priced for a scenario that isn't going to play out? Note: Baker runs Atreides Management, one of the most respected tech-focused growth funds in the world. He doesn't make calls like this lightly. When the best investors in tech spot a cross-sectional mismatch this size, the trade is figuring out which side wins. Our analysts are already positioning inside the AI supply chain. They called $AMD, $MU, $CRDO and $NBIS before the big runs. Follow their exact portfolios for $1 at Milk Road PRO, link in bio.

Milk Road Macro

162,479 次观看 • 2 个月前

New course to bring you up to state-of-the-art at using AI to help you code: Build Apps with Windsurf's AI Coding Agents, built in partnership with WIndsurf (Codeium) and taught by Anshul Ramachandran! AI-assisted IDEs (Integrated Development Environments) make developers’ workflows faster, more efficient, and much more fun. Agentic tools like Windsurf are more than just code autocomplete—they are collaborative coding agents that help you break down complex applications, iterate efficiently, and generate code that spans multiple files. Although a lot of coding assistants share the same underlying large language models for planning and reasoning, a major point of distinction is how they handle tools, keep track of context, and stay aligned with your intent as a developer. For instance, if you make modifications to a class definition in your code and make the same modifications to other classes in the same directory, you might tell the AI agent "Do the same thing in similar places in this directory." Here, tracking your intent means understanding that “the same thing" refers to that recent edit you just made, which must be followed by appropriate search and tool-calling to implement the changes. In this course, you'll learn the inner workings of coding agents, their strengths and limitations, and how to use Windsurf to quickly build several applications. In detail, you'll: - Build a mental model of how agents work by combining human-action tracking, tool integration, and context awareness to carry out an agentic coding workflow. - Learn the challenges of code search and discovery and how a multi-step retrieval approach helps coding agents address them. - Use Windsurf to analyze and understand a large, old codebase and update it to the latest versions of the frameworks and packages it uses. - Build a Wikipedia data analysis app that retrieves, parses, and analyzes word frequencies. - Enhance the performance of your Wikipedia analysis app by adding caching, and through this, also learn how to course-correct when the AI agent produces unexpected results. - Learn tips and tricks such as keyboard shortcuts, autocomplete, and @ mentions to quickly call on agentic capabilities. - Use image/multimodal capabilities of the AI agent to increase your development velocity; you'll see an example of uploading a mockup with sketched-out UI features, and ask the agent to use that to build new functionality to an app. By the end of this course, you’ll understand agentic coding in-depth and know how to use it to make your development process much faster, more efficient, and enjoyable. Please sign up here!

New course to bring you up to state-of-the-art at using AI to help you code: Build Apps with Windsurf's AI Coding Agents, built in partnership with WIndsurf (Codeium) and taught by Anshul Ramachandran! AI-assisted IDEs (Integrated Development Environments) make developers’ workflows faster, more efficient, and much more fun. Agentic tools like Windsurf are more than just code autocomplete—they are collaborative coding agents that help you break down complex applications, iterate efficiently, and generate code that spans multiple files. Although a lot of coding assistants share the same underlying large language models for planning and reasoning, a major point of distinction is how they handle tools, keep track of context, and stay aligned with your intent as a developer. For instance, if you make modifications to a class definition in your code and make the same modifications to other classes in the same directory, you might tell the AI agent "Do the same thing in similar places in this directory." Here, tracking your intent means understanding that “the same thing" refers to that recent edit you just made, which must be followed by appropriate search and tool-calling to implement the changes. In this course, you'll learn the inner workings of coding agents, their strengths and limitations, and how to use Windsurf to quickly build several applications. In detail, you'll: - Build a mental model of how agents work by combining human-action tracking, tool integration, and context awareness to carry out an agentic coding workflow. - Learn the challenges of code search and discovery and how a multi-step retrieval approach helps coding agents address them. - Use Windsurf to analyze and understand a large, old codebase and update it to the latest versions of the frameworks and packages it uses. - Build a Wikipedia data analysis app that retrieves, parses, and analyzes word frequencies. - Enhance the performance of your Wikipedia analysis app by adding caching, and through this, also learn how to course-correct when the AI agent produces unexpected results. - Learn tips and tricks such as keyboard shortcuts, autocomplete, and @ mentions to quickly call on agentic capabilities. - Use image/multimodal capabilities of the AI agent to increase your development velocity; you'll see an example of uploading a mockup with sketched-out UI features, and ask the agent to use that to build new functionality to an app. By the end of this course, you’ll understand agentic coding in-depth and know how to use it to make your development process much faster, more efficient, and enjoyable. Please sign up here!

Andrew Ng

139,873 次观看 • 1 年前

I designed a new test specifically for multimodal models: fill out a paper form. And it's much harder than it sounds. This isn't typing into an electronic field that captures your text. The form is just an image. The model has to place each form element: text, checkmarks — at the correct pixel position on the canvas itself. Results: 🟢 Kimi K2.6 → done in 3:45, 16.7k output tokens 🟡 Step 3.7 Flash → half the fields, 57k output tokens 🔴 Gemini 3.5 Flash → 489k output tokens, never finished. I had to kill it. Gemini burned ~29x more output tokens than Kimi on the exact same task, and Kimi's was the only form that actually looked filled out. The test, a mocked application form, contains some challenging parts, such as one-character-per-box fields. I provided every model the same set of tools: > get canvas size > drop probe markers to find coordinates > add text > add checkmarks > move elements > take a screenshot anytime to check their own work > ... etc So it's vision + spatial reasoning + tool use + long context, all at once. Small models (Qwen, Gemma) can't really complete this test, so I skipped them. What happened: > Kimi nailed name, DOB, ID, gender, marital status, nationality, email, phone, address, postal code — placement slightly loose, but content correct. 15 turns. Clean. > Step got maybe half right — fields dropped, "United States" landed in the email line, data floating outside boxes. Burned 1.24M input tokens doing it (81 turns of re-reading the canvas). > Gemini almost got there visually... then spiraled. By turn 40 it was issuing a delete_elements call wiping element IDs 365–425, basically erasing its own work. 31 minutes, 489k output tokens, still streaming. Terminated. The takeaway isn't "Gemini bad." This test is indeed difficult. But token efficiency is capability now. A model that needs 30x the tokens and still can't converge is going to be 30x the cost in production. Kimi K2.6 just quietly did the thing.

I designed a new test specifically for multimodal models: fill out a paper form. And it's much harder than it sounds. This isn't typing into an electronic field that captures your text. The form is just an image. The model has to place each form element: text, checkmarks — at the correct pixel position on the canvas itself. Results: 🟢 Kimi K2.6 → done in 3:45, 16.7k output tokens 🟡 Step 3.7 Flash → half the fields, 57k output tokens 🔴 Gemini 3.5 Flash → 489k output tokens, never finished. I had to kill it. Gemini burned ~29x more output tokens than Kimi on the exact same task, and Kimi's was the only form that actually looked filled out. The test, a mocked application form, contains some challenging parts, such as one-character-per-box fields. I provided every model the same set of tools: > get canvas size > drop probe markers to find coordinates > add text > add checkmarks > move elements > take a screenshot anytime to check their own work > ... etc So it's vision + spatial reasoning + tool use + long context, all at once. Small models (Qwen, Gemma) can't really complete this test, so I skipped them. What happened: > Kimi nailed name, DOB, ID, gender, marital status, nationality, email, phone, address, postal code — placement slightly loose, but content correct. 15 turns. Clean. > Step got maybe half right — fields dropped, "United States" landed in the email line, data floating outside boxes. Burned 1.24M input tokens doing it (81 turns of re-reading the canvas). > Gemini almost got there visually... then spiraled. By turn 40 it was issuing a delete_elements call wiping element IDs 365–425, basically erasing its own work. 31 minutes, 489k output tokens, still streaming. Terminated. The takeaway isn't "Gemini bad." This test is indeed difficult. But token efficiency is capability now. A model that needs 30x the tokens and still can't converge is going to be 30x the cost in production. Kimi K2.6 just quietly did the thing.

stevibe

25,446 次观看 • 2 个月前

I deleted half of my AI-video prompt — and the result got better. 😶 Here's the before/after that changed how I prompt, and the skill I open-sourced from it. 🧵 We've been prompting video models like a film shot list: 24mm, f/1.4, "volumetric fluid simulation," frame-by-frame timing. But a lot of the time you don't need to — the smarter the model gets, the simpler the prompt can be: tell the story, the texture of the air, the emotion, and let it pick the shots, light, and rhythm. ByteDance shipped this idea with Seedance 2.0 and called it "Vibe Creating" — I open-sourced the skill and the philosophy behind it. Same scene, two prompts 👇 🔧 Regular: "85mm f1.4 macro, 120fps, dolly 0.6x, freeze the sweat at 1/250…" ✨ Vibe: "Late-night street stall. The cook flicks the wok; a ball of orange flame lights up his sweating face. Noodles fly. He plates them and wipes his brow." Same model. One of them feels alive. The skill is built for story-driven video — concept shorts, micro-narratives, emotional or atmosphere pieces, anything where you'd rather describe the moment than dictate the camera. Feed it your idea (or your over-stuffed prompt) and it hands back the version the model actually shoots better. The one place it won't go: UI demos, step-by-step tutorials, or exact dialogue sync — there it'll tell you it's the wrong tool rather than flatten a prompt that needed to stay precise. Try it on your next prompt — repo 👇

I deleted half of my AI-video prompt — and the result got better. 😶 Here's the before/after that changed how I prompt, and the skill I open-sourced from it. 🧵 We've been prompting video models like a film shot list: 24mm, f/1.4, "volumetric fluid simulation," frame-by-frame timing. But a lot of the time you don't need to — the smarter the model gets, the simpler the prompt can be: tell the story, the texture of the air, the emotion, and let it pick the shots, light, and rhythm. ByteDance shipped this idea with Seedance 2.0 and called it "Vibe Creating" — I open-sourced the skill and the philosophy behind it. Same scene, two prompts 👇 🔧 Regular: "85mm f1.4 macro, 120fps, dolly 0.6x, freeze the sweat at 1/250…" ✨ Vibe: "Late-night street stall. The cook flicks the wok; a ball of orange flame lights up his sweating face. Noodles fly. He plates them and wipes his brow." Same model. One of them feels alive. The skill is built for story-driven video — concept shorts, micro-narratives, emotional or atmosphere pieces, anything where you'd rather describe the moment than dictate the camera. Feed it your idea (or your over-stuffed prompt) and it hands back the version the model actually shoots better. The one place it won't go: UI demos, step-by-step tutorials, or exact dialogue sync — there it'll tell you it's the wrong tool rather than flatten a prompt that needed to stay precise. Try it on your next prompt — repo 👇

Alisa Qian

11,983 次观看 • 1 个月前

R.I.P. dev agencies ☠️ There's now an AI that debugs your app, clicks through the UI, finds what's broken, and fixes it on its own. And it hits a 97% success rate. Max isn’t a “code assistant.” It behaves like a full software engineer that never gets tired, never drifts, and never waits for your next instruction. You give it a goal. Max writes the code, tests it, checks the UI, finds the issues, fixes them, and loops until the task is done. Not theory. Actual work. Max handled things that normally eat entire days: • Debugging a Stripe checkout that refused to load • Solving a login-on-refresh issue by tracking network behavior step by step • Building a full Teams & Invites feature UI, routes, tokens, emails, tests while I focused on another part of the product 97 percent of the problems I threw at it got solved without me stepping in. The wild part is how Max operates: It sees the app. It clicks around. It interacts with the environment. It rewrites code. It retries until it works. Goal → attempt → improve → retry → done. Some tasks take 100 moves. Some run for 30 minutes. Most finish before I even finish a coffee. This flips the entire dev workflow on its head. Instead of “write code and hope it works,” you define outcomes and watch the engineer in the background get it done. Every once in a while a tool shows up that changes the tempo of an entire industry. Anything Max is one of them.

R.I.P. dev agencies ☠️ There's now an AI that debugs your app, clicks through the UI, finds what's broken, and fixes it on its own. And it hits a 97% success rate. Max isn’t a “code assistant.” It behaves like a full software engineer that never gets tired, never drifts, and never waits for your next instruction. You give it a goal. Max writes the code, tests it, checks the UI, finds the issues, fixes them, and loops until the task is done. Not theory. Actual work. Max handled things that normally eat entire days: • Debugging a Stripe checkout that refused to load • Solving a login-on-refresh issue by tracking network behavior step by step • Building a full Teams & Invites feature UI, routes, tokens, emails, tests while I focused on another part of the product 97 percent of the problems I threw at it got solved without me stepping in. The wild part is how Max operates: It sees the app. It clicks around. It interacts with the environment. It rewrites code. It retries until it works. Goal → attempt → improve → retry → done. Some tasks take 100 moves. Some run for 30 minutes. Most finish before I even finish a coffee. This flips the entire dev workflow on its head. Instead of “write code and hope it works,” you define outcomes and watch the engineer in the background get it done. Every once in a while a tool shows up that changes the tempo of an entire industry. Anything Max is one of them.

Alex Veremeyenko

80,509 次观看 • 8 个月前

【Wouldn’t Having a Performance Engineer from Formula E Bring an Advantage?】 Sky Germany asked Toto whether Ant had an advantage because his side of the garage has a Performance Engineer with a Formula E background. Toto dismissed the idea, saying both sides have complete access to each other’s data and tools. But while data and tools can be shared, can experience, judgment, and modeling expertise really be shared as well? That engineer is Stéphane Folio, who previously worked as a Performance Engineer in Nissan Formula E, specializing in energy management, chassis control strategy, vehicle control strategy, the Driver-in-the-Loop simulator, and vehicle dynamics. He joined Mercedes F1 in 2025—the same year as Ant. Why is Formula E experience so valuable under the 2026 regulations? Because after the first ten races, we have already seen that a significant part of a car’s performance is determined by when energy is harvested, when it is deployed, how the state of charge is managed, and how the software calculates and controls that deployment. This may also explain why the gaps between some teammates who were far apart last season have narrowed significantly this year—or even been reversed. Formula E uses a spec battery and several standardized core components, meaning teams cannot rely on hardware alone to create a major advantage. As a result, performance is heavily centered on powertrain efficiency, energy management, control strategies, and software calibration. Engineers spend years optimizing exactly how much energy to recover in each corner, how much to deploy on every straight, and how to recalculate the entire energy strategy after safety cars, slipstreaming, or changes in race position. In many ways, Formula E is a championship where performance is heavily defined by software and energy strategy—experience that overlaps remarkably well with the key technical challenges of the 2026 F1 regulations. Mercedes itself also competed as a factory team in Formula E, winning two Drivers’ Championships, one Teams’ Championship, and one Teams’ runners-up finish. So wouldn’t assigning an engineer with exactly this kind of background to Ant’s side of the garage be a resource advantage in itself? Toto says both sides can see the same data. But seeing the same data does not mean having the same people, the same experience, or the ability to reach equally fast and equally accurate conclusions. Meanwhile, in nine of the first ten races this season, GR has suffered different software or hardware-related issues.

【Wouldn’t Having a Performance Engineer from Formula E Bring an Advantage?】 Sky Germany asked Toto whether Ant had an advantage because his side of the garage has a Performance Engineer with a Formula E background. Toto dismissed the idea, saying both sides have complete access to each other’s data and tools. But while data and tools can be shared, can experience, judgment, and modeling expertise really be shared as well? That engineer is Stéphane Folio, who previously worked as a Performance Engineer in Nissan Formula E, specializing in energy management, chassis control strategy, vehicle control strategy, the Driver-in-the-Loop simulator, and vehicle dynamics. He joined Mercedes F1 in 2025—the same year as Ant. Why is Formula E experience so valuable under the 2026 regulations? Because after the first ten races, we have already seen that a significant part of a car’s performance is determined by when energy is harvested, when it is deployed, how the state of charge is managed, and how the software calculates and controls that deployment. This may also explain why the gaps between some teammates who were far apart last season have narrowed significantly this year—or even been reversed. Formula E uses a spec battery and several standardized core components, meaning teams cannot rely on hardware alone to create a major advantage. As a result, performance is heavily centered on powertrain efficiency, energy management, control strategies, and software calibration. Engineers spend years optimizing exactly how much energy to recover in each corner, how much to deploy on every straight, and how to recalculate the entire energy strategy after safety cars, slipstreaming, or changes in race position. In many ways, Formula E is a championship where performance is heavily defined by software and energy strategy—experience that overlaps remarkably well with the key technical challenges of the 2026 F1 regulations. Mercedes itself also competed as a factory team in Formula E, winning two Drivers’ Championships, one Teams’ Championship, and one Teams’ runners-up finish. So wouldn’t assigning an engineer with exactly this kind of background to Ant’s side of the garage be a resource advantage in itself? Toto says both sides can see the same data. But seeing the same data does not mean having the same people, the same experience, or the ability to reach equally fast and equally accurate conclusions. Meanwhile, in nine of the first ten races this season, GR has suffered different software or hardware-related issues.

Holy_Moly

55,499 次观看 • 17 天前

Atomic Agent beat Hermes on GAIA: 69.8% vs 58.5%, and it was 1.6x faster! We ran both agents through the full GAIA Level 1 benchmark, 53 real-world tasks, same 4-bit qwen-3.6-35b on the same Apple M4 Max. Results: ✦ Atomic Agent: 37 of 53 solved, done in 3h 12m ✦ Hermes Agent: 31 of 53 solved, took 5h 10m Atomic solved 6 more tasks and finished nearly 2 hours sooner. Hermes ran into the 900s timeout on 7 tasks; Atomic on just 2. Hermes burned 71% of its total time on tasks it still failed, Atomic, 48%. Where it showed: ✦ Audre Lorde poem, which stanza is indented: Atomic pushed through a dead source, switched tools, and answered in 7.6 min. Hermes ran the full clock and returned a blank. ✦ Vietnamese specimens, which city they ended up in: Atomic pulled it from the first source and normalized the answer in 33s. Hermes spent 7.3 min and never answered. ✦ The dinosaur featured-article nominator: Atomic walked the Wikipedia chain to "FunkMonk" in 57s. Hermes guessed a wrong name after 11 min. Atomic keeps a byte-stable prompt prefix, so llama-server reuses the KV-cache instead of re-encoding the whole context every turn, and it emits one JSON array of tool calls per inference, then compresses results back instead of pasting them in full, so the context never balloons and a small model stays sharp deep into a task. On top of that a no-progress guard vetoes repeated identical tool calls (warn at 3, hard veto at 5) and forces a reply, so Atomic never sinks 15 minutes into re-scanning one page the way Hermes did. Both agents missed some of the same questions, and on a few Hermes got there and Atomic did not, usually format slips where Atomic computed the right number but printed the working instead of the bare value. But on identical hardware and identical weights, the runtime that reuses its cache and refuses to spin came out ahead on accuracy and speed. Getting this from the runtime alone is wild. Run the same 53 GAIA tasks on Atomic Agent!

Atomic Agent beat Hermes on GAIA: 69.8% vs 58.5%, and it was 1.6x faster! We ran both agents through the full GAIA Level 1 benchmark, 53 real-world tasks, same 4-bit qwen-3.6-35b on the same Apple M4 Max. Results: ✦ Atomic Agent: 37 of 53 solved, done in 3h 12m ✦ Hermes Agent: 31 of 53 solved, took 5h 10m Atomic solved 6 more tasks and finished nearly 2 hours sooner. Hermes ran into the 900s timeout on 7 tasks; Atomic on just 2. Hermes burned 71% of its total time on tasks it still failed, Atomic, 48%. Where it showed: ✦ Audre Lorde poem, which stanza is indented: Atomic pushed through a dead source, switched tools, and answered in 7.6 min. Hermes ran the full clock and returned a blank. ✦ Vietnamese specimens, which city they ended up in: Atomic pulled it from the first source and normalized the answer in 33s. Hermes spent 7.3 min and never answered. ✦ The dinosaur featured-article nominator: Atomic walked the Wikipedia chain to "FunkMonk" in 57s. Hermes guessed a wrong name after 11 min. Atomic keeps a byte-stable prompt prefix, so llama-server reuses the KV-cache instead of re-encoding the whole context every turn, and it emits one JSON array of tool calls per inference, then compresses results back instead of pasting them in full, so the context never balloons and a small model stays sharp deep into a task. On top of that a no-progress guard vetoes repeated identical tool calls (warn at 3, hard veto at 5) and forces a reply, so Atomic never sinks 15 minutes into re-scanning one page the way Hermes did. Both agents missed some of the same questions, and on a few Hermes got there and Atomic did not, usually format slips where Atomic computed the right number but printed the working instead of the bare value. But on identical hardware and identical weights, the runtime that reuses its cache and refuses to spin came out ahead on accuracy and speed. Getting this from the runtime alone is wild. Run the same 53 GAIA tasks on Atomic Agent!

Atomic Agent

111,357 次观看 • 17 天前

To every NEET candidate and parent - please watch this. A few minutes that could save you serious money and stress this week. Scammers on Telegram were running two rackets targeting you: 1️⃣ Channels demanding ₹14,000 to ₹25,000 - some even ₹10 lakh - claiming they'll send you the re-exam paper. They won't. There is no leaked paper for the re-exam. The money is gone the moment you transfer it. Your admit card and WhatsApp number, if you sent them, become the tools they use to scam the next student. 2️⃣ Fake "proof" videos showing chats from before the exam. The trick: on Telegram, whoever runs a channel can edit any old message AND change what's inside it, while the date on the message stays the same. So a message edited on the 4th can be made to look exactly like it was sent on the 1st. The full explanation is in the video below — please watch and share with anyone you know who's preparing 👇 🎥 The same trick will be tried again after 21st June. Don't fall for it. Don't forward it. Don't pay anyone. ✅ Focus on your prep - you've earned this ✅ Trust only and verified NTA handles ✅ Tell your friends - especially anyone anxious enough to be tempted 📞 Report any scam: National Cyber-Crime Helpline 1930 or Your hard work is what will get you through this exam. Not a Telegram channel. We're on your side. 💪 🎥 Prof. V. Kamakoti, Director, IIT Madras, also explains the technical side clearly - his videos are in the replies below.

To every NEET candidate and parent - please watch this. A few minutes that could save you serious money and stress this week. Scammers on Telegram were running two rackets targeting you: 1️⃣ Channels demanding ₹14,000 to ₹25,000 - some even ₹10 lakh - claiming they'll send you the re-exam paper. They won't. There is no leaked paper for the re-exam. The money is gone the moment you transfer it. Your admit card and WhatsApp number, if you sent them, become the tools they use to scam the next student. 2️⃣ Fake "proof" videos showing chats from before the exam. The trick: on Telegram, whoever runs a channel can edit any old message AND change what's inside it, while the date on the message stays the same. So a message edited on the 4th can be made to look exactly like it was sent on the 1st. The full explanation is in the video below — please watch and share with anyone you know who's preparing 👇 🎥 The same trick will be tried again after 21st June. Don't fall for it. Don't forward it. Don't pay anyone. ✅ Focus on your prep - you've earned this ✅ Trust only and verified NTA handles ✅ Tell your friends - especially anyone anxious enough to be tempted 📞 Report any scam: National Cyber-Crime Helpline 1930 or Your hard work is what will get you through this exam. Not a Telegram channel. We're on your side. 💪 🎥 Prof. V. Kamakoti, Director, IIT Madras, also explains the technical side clearly - his videos are in the replies below.

National Testing Agency

771,852 次观看 • 1 个月前

10 repos that cut your ai agent token bill by up to 80% 1. microsoft/LLMLingua → cuts prompt size by up to 95% compresses prompts before the api call. 20x compression. published at EMNLP + ACL. near-zero quality loss. 6,100 stars 2. mem0ai/mem0 → replaces full conversation history in context stores what matters. retrieves only what's needed. 10,000 token history → 200 token memory. per agent. 54,800 stars 3. BerriAI/litellm → routes each call to the cheapest model simple task → haiku. complex task → sonnet. tracks cost per agent, per call, per day. 45,700 stars 4. run-llama/llama_index → replaces sending full documents rag: 100-page doc → 3 relevant chunks → same answer. 98% fewer tokens per query. 49,100 stars 5. chroma-core/chroma → replaces keyword search in full context vector store. finds the closest match. feeds only that. 50-200 tokens per query instead of thousands. 27,800 stars 6. letta-ai/letta → replaces infinite context window crashes paged memory for agents. loads only relevant memory. stops your agent from hitting limits and retrying. 22,400 stars 7. guidance-ai/guidance → cuts output token bloat by 30-50% structured generation. constrains model output natively. no more 100-token prompts to get json back. 21,400 stars 8. Aider-AI/aider → replaces pasting entire codebases builds a repo map. sends only files relevant to the task. not your whole project. just what the agent needs. 44,300 stars 9. openai/tiktoken → count tokens before you send know the exact cost before the api call happens. not after the bill arrives. 18,100 stars 10. simonw/ttok → hard cap on what gets sent cli tool: count tokens, truncate to budget limit. pipe any text in. get truncated output back. 389 stars most agents are expensive not because the model is expensive. because nobody checked what was being sent to it.

10 repos that cut your ai agent token bill by up to 80% 1. microsoft/LLMLingua → cuts prompt size by up to 95% compresses prompts before the api call. 20x compression. published at EMNLP + ACL. near-zero quality loss. 6,100 stars 2. mem0ai/mem0 → replaces full conversation history in context stores what matters. retrieves only what's needed. 10,000 token history → 200 token memory. per agent. 54,800 stars 3. BerriAI/litellm → routes each call to the cheapest model simple task → haiku. complex task → sonnet. tracks cost per agent, per call, per day. 45,700 stars 4. run-llama/llama_index → replaces sending full documents rag: 100-page doc → 3 relevant chunks → same answer. 98% fewer tokens per query. 49,100 stars 5. chroma-core/chroma → replaces keyword search in full context vector store. finds the closest match. feeds only that. 50-200 tokens per query instead of thousands. 27,800 stars 6. letta-ai/letta → replaces infinite context window crashes paged memory for agents. loads only relevant memory. stops your agent from hitting limits and retrying. 22,400 stars 7. guidance-ai/guidance → cuts output token bloat by 30-50% structured generation. constrains model output natively. no more 100-token prompts to get json back. 21,400 stars 8. Aider-AI/aider → replaces pasting entire codebases builds a repo map. sends only files relevant to the task. not your whole project. just what the agent needs. 44,300 stars 9. openai/tiktoken → count tokens before you send know the exact cost before the api call happens. not after the bill arrives. 18,100 stars 10. simonw/ttok → hard cap on what gets sent cli tool: count tokens, truncate to budget limit. pipe any text in. get truncated output back. 389 stars most agents are expensive not because the model is expensive. because nobody checked what was being sent to it.

self.dll

39,554 次观看 • 3 个月前