Загрузка видео...

Не удалось загрузить видео

На главную

Mood: agents optimizing kernels Claude won on kernel optimization: gemm_bf16 at 1.19x vs Codex's 0.94x. Codex was faster (~1.3h vs ~3.4h) but produced no reinjectable optimizations. Claude used (hipBLASLt) as a drop-in replacement for the custom Triton kernel. For Codex, shape mismatch caused slight regression. Still improving, open sourcing...

realSharonZhou's profile picture

Sharon Zhou

27,371 subscribers

23,583 просмотров • 4 месяцев назад •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

Комментарии: 0

Нет доступных комментариев

Здесь появятся комментарии из оригинального поста

Похожие видео

Some time ago, I had the idea to port NVIDIA Physical AI stack to AMD. The motivation was to improve hardware diversity and enable world models and VLAs to run beyond a single ecosystem. We started with NVIDIA Cosmos Predict 2.5-2B. Porting wasn’t trivial: these models are deeply optimized for NVIDIA’s stack. We used this as an opportunity to apply our ROCm kernels. The results were surprising: Both encode and diffusion run faster on AMD Instinct MI300X vs. NVIDIA H200 (FA3) and we still saw significant headroom for further optimization. Quality is unchanged across modalities (validated with WorldJen) To be clear, this is no luck. We have deep experience with diffusion models and AMD GPUs. But this just gives us a good opportunity to get closer to a true hardware-to-hardware comparison, as we work with less software abstractions than usual. Just to give an example, on AMD, memory instructions are async with a hardware queue of ordered pending instructions, enabling concurrent load/store with compute without warp specialization. Bottom line: there are real architectural advantages on AMD, if you take the time to work with the hardware. Note, we did tradeoff ~20% higher memory usage, That being said, AMD has more to give to begin with :) in the coming weeks: AMD versions of Cosmos Transfer and GR00T, an even faster version of Cosmos Predict, and open-sourcing an attention kernel faster than AITER v3 (which is closed-source for some reason? cc: Anush Elangovan )
0:44

Some time ago, I had the idea to port NVIDIA Physical AI stack to AMD. The motivation was to improve hardware diversity and enable world models and VLAs to run beyond a single ecosystem. We started with NVIDIA Cosmos Predict 2.5-2B. Porting wasn’t trivial: these models are deeply optimized for NVIDIA’s stack. We used this as an opportunity to apply our ROCm kernels. The results were surprising: Both encode and diffusion run faster on AMD Instinct MI300X vs. NVIDIA H200 (FA3) and we still saw significant headroom for further optimization. Quality is unchanged across modalities (validated with WorldJen) To be clear, this is no luck. We have deep experience with diffusion models and AMD GPUs. But this just gives us a good opportunity to get closer to a true hardware-to-hardware comparison, as we work with less software abstractions than usual. Just to give an example, on AMD, memory instructions are async with a hardware queue of ordered pending instructions, enabling concurrent load/store with compute without warp specialization. Bottom line: there are real architectural advantages on AMD, if you take the time to work with the hardware. Note, we did tradeoff ~20% higher memory usage, That being said, AMD has more to give to begin with :) in the coming weeks: AMD versions of Cosmos Transfer and GR00T, an even faster version of Cosmos Predict, and open-sourcing an attention kernel faster than AITER v3 (which is closed-source for some reason? cc: Anush Elangovan )

Omer Shlomovits

36,620 просмотров • 3 месяцев назад

Three skills I use every day in Claude Code and Codex to solve my hardest problems: 1️⃣ /agent-watchdog When I have one agent like Codex working on a task and I don't fully trust it's going to do everything right, I'll open up another one like Claude Code and tell it to watchdog the Codex thread. You can copy the Codex deep link into Claude Code and it'll look at the prompt you sent, watch the Codex thread until it's done, then compare the Codex solution to how it was planning to solve it and automatically fix anything that Codex missed. It can also test the work of the other agent end-to-end. Similar to the idea of OpenRouter's new Fusion feature, I've definitely found that two models thinking through a problem and checking each other's work can be wildly more impactful than just one. 2️⃣ /plan-arbiter Similar ideas as /agent-watchdog - but with this one you have both make plans, compare plans, negotiate the differences, and make a final plan to execute. I find Claude Code is better at writing plans, but Codex is faster and cheaper to execute on them. Then I usually have Claude Code watchdog the Codex work and fix anything that was missed. 3️⃣ /read-the-damn-docs One thing that drives me crazy with coding agents is they're so reluctant to look up docs. They'll just guess and guess and guess at the right API surface for things, or the right solution to an integration of two things. Once I explicitly tell it to look up the docs, it says "Oh, I see the answer," and it fixes the problem. So I made the /read-the-damn-docs skill. Add it and your agents will know when and how to do efficient web searches to look up docs for the types of problems you really should look up docs for. All of these are totally open source over on my GitHub. If you try them, let me know your feedback. Will link to them below:
1:21

Three skills I use every day in Claude Code and Codex to solve my hardest problems: 1️⃣ /agent-watchdog When I have one agent like Codex working on a task and I don't fully trust it's going to do everything right, I'll open up another one like Claude Code and tell it to watchdog the Codex thread. You can copy the Codex deep link into Claude Code and it'll look at the prompt you sent, watch the Codex thread until it's done, then compare the Codex solution to how it was planning to solve it and automatically fix anything that Codex missed. It can also test the work of the other agent end-to-end. Similar to the idea of OpenRouter's new Fusion feature, I've definitely found that two models thinking through a problem and checking each other's work can be wildly more impactful than just one. 2️⃣ /plan-arbiter Similar ideas as /agent-watchdog - but with this one you have both make plans, compare plans, negotiate the differences, and make a final plan to execute. I find Claude Code is better at writing plans, but Codex is faster and cheaper to execute on them. Then I usually have Claude Code watchdog the Codex work and fix anything that was missed. 3️⃣ /read-the-damn-docs One thing that drives me crazy with coding agents is they're so reluctant to look up docs. They'll just guess and guess and guess at the right API surface for things, or the right solution to an integration of two things. Once I explicitly tell it to look up the docs, it says "Oh, I see the answer," and it fixes the problem. So I made the /read-the-damn-docs skill. Add it and your agents will know when and how to do efficient web searches to look up docs for the types of problems you really should look up docs for. All of these are totally open source over on my GitHub. If you try them, let me know your feedback. Will link to them below:

Steve (Builder.io)

42,501 просмотров • 1 месяц назад

I just compared Claude Code vs Codex vs Cursor CLI The task was to build a Next.js app with Tailwind 4 and shadcn components to collect customer feedback and showcase it with a widget. I gave all three the same prompt and let them go for 30 minutes to see what they came up with. Claude Code with Opus 4.1 Even though I told it to set up the app in the existing project folder, it tried to create a directory for it. After I interrupted and told it not to do that, it built a demo form and landing page with no errors. I had to ask it to make the demo interactive so users could submit a testimonial and preview it. The landing page looked like AI and was pretty basic, but it worked and it was done in a fraction of the time of the others. Total tokens used: 33k Codex with GPT-5 At the end of the 30 minutes I just could not get Codex to produce a working app. It got stuck in a loop of not being able to set up Tailwind 4 and despite many, MANY, attempts, I ended up with a "failed to compile" error. Total tokens used: 102k Cursor Agent with GPT-5 This was the slowest agent by far and a couple of times I actually thought it got stuck in a loop and was close to Ctrl+C'ing to cancel it. The TUI is really nice though, especially how it shows diffs and it did eventually build a working app (after one or two slight errors that needed fixing) The demo was interactive and it had a very minimal design that looked bare but also a lot less like an "AI generated" app than the Opus 4.1 design. It also wasn't too chatty and just did what it needed to do! Code quality was on a par with Opus 4.1, but it did use 5.5x as many tokens to get there. Still cheaper than Opus on a direct comparison but not when you factor in a Claude Code Max subscription. Total tokens: 188k I'll be able to do a proper comparison and record some videos when I'm back from holiday but for now, Opus is still the more capable model out of the box and Claude Code is the more complete CLI product. It will be interesting to see how Cursor evolve their CLI though with commands and subagents because I think with GPT-5 they have a real shot at providing competition for Claude Code if they can optimise output to get similar quality with less tokens. Jump to 0:40 in the video to see the two apps. Which do you think is which? ;)
1:18

I just compared Claude Code vs Codex vs Cursor CLI The task was to build a Next.js app with Tailwind 4 and shadcn components to collect customer feedback and showcase it with a widget. I gave all three the same prompt and let them go for 30 minutes to see what they came up with. Claude Code with Opus 4.1 Even though I told it to set up the app in the existing project folder, it tried to create a directory for it. After I interrupted and told it not to do that, it built a demo form and landing page with no errors. I had to ask it to make the demo interactive so users could submit a testimonial and preview it. The landing page looked like AI and was pretty basic, but it worked and it was done in a fraction of the time of the others. Total tokens used: 33k Codex with GPT-5 At the end of the 30 minutes I just could not get Codex to produce a working app. It got stuck in a loop of not being able to set up Tailwind 4 and despite many, MANY, attempts, I ended up with a "failed to compile" error. Total tokens used: 102k Cursor Agent with GPT-5 This was the slowest agent by far and a couple of times I actually thought it got stuck in a loop and was close to Ctrl+C'ing to cancel it. The TUI is really nice though, especially how it shows diffs and it did eventually build a working app (after one or two slight errors that needed fixing) The demo was interactive and it had a very minimal design that looked bare but also a lot less like an "AI generated" app than the Opus 4.1 design. It also wasn't too chatty and just did what it needed to do! Code quality was on a par with Opus 4.1, but it did use 5.5x as many tokens to get there. Still cheaper than Opus on a direct comparison but not when you factor in a Claude Code Max subscription. Total tokens: 188k I'll be able to do a proper comparison and record some videos when I'm back from holiday but for now, Opus is still the more capable model out of the box and Claude Code is the more complete CLI product. It will be interesting to see how Cursor evolve their CLI though with commands and subagents because I think with GPT-5 they have a real shot at providing competition for Claude Code if they can optimise output to get similar quality with less tokens. Jump to 0:40 in the video to see the two apps. Which do you think is which? ;)

Ian Nuttall

194,949 просмотров • 11 месяцев назад

OpenAI’s hottest app isn’t ChatGPT—it’s Codex. In the last few weeks alone, the Codex team shipped a desktop app, GPT-5.3 Codex (a new flagship model), and Spark, the fastest coding model I’ve ever used. Usage has grown fivefold since January and over a million people now use Codex weekly. Codex was also the app that OpenAI chose to run an ad for in the Super Bowl. I talked to Thibault (Tibo), head of Codex, and Andrew (Andrew Ambrosino), a member of technical staff who built the Codex app, for Every 📧’s AI & I about what OpenAI is building and how they’re using it internally. We get into: - Why they built a GUI instead of a terminal. Terminals work for quick tasks, they say, but feel limiting when you’re running multiple agents in parallel. The IDE, meanwhile, overwhelms users—and the Codex team wants the AI to dynamically decide which tools to show you for a given task. - How they’re teaching the model to read between the lines. Codex is great at following instructions, but optimize too hard in that direction, and it starts taking you literally—like copying a typo directly into the code. The team obsesses over this tradeoff, and is also introducing “personalities,” modes users can toggle between that control how blunt or supportive the model feels. - How OpenAI uses its own coding agent. Codex lets you schedule prompts to run on a recurring basis, and the team has dozens of automations running at all times. For example, one scans for merge conflicts every couple of hours so code is always ready to ship, and another picks a random file from the codebase multiple times a day and hunts for bugs no one would've gone looking for. - Why speed is a dimension of intelligence. OpenAI’s newest model (Spark) is so fast that they actually slow it down so you can read the output. They see the speed enabling three things: staying super in the flow, replacing brittle developer tools with intelligent ones that can adapt on the fly, and redirecting the model mid-task— especially with voice—so coding starts to feel more and more like a conversation. - Code review is the next bottleneck. Models can generate code faster than ever, but someone still has to verify that it works. The team is exploring a future where the model proves its own fix works—retracing the click path a user would take, screenshotting the results, and attaching the evidence to a pull request. This is a must-watch for anyone who uses AI coding agents—and is curious about the future of programming. Watch below! Timestamps: Introduction: 00:01:27 OpenAI’s evolving bet on its coding agent: 00:05:27 The choice to invest in a GUI (over a terminal): 00:09:42 The AI workflows that the Codex team relies on to ship: 00:20:38 Teaching Codex how to read between the lines: 00:26:45 Building affordances for a lightening fast model: 00:28:45 Why speed is a dimension of intelligence: 00:33:15 Code review is the next bottleneck for coding agents: 00:36:30 How the Codex team positions against the competition: 00:41:24
46:40

OpenAI’s hottest app isn’t ChatGPT—it’s Codex. In the last few weeks alone, the Codex team shipped a desktop app, GPT-5.3 Codex (a new flagship model), and Spark, the fastest coding model I’ve ever used. Usage has grown fivefold since January and over a million people now use Codex weekly. Codex was also the app that OpenAI chose to run an ad for in the Super Bowl. I talked to Thibault (Tibo), head of Codex, and Andrew (Andrew Ambrosino), a member of technical staff who built the Codex app, for Every 📧’s AI & I about what OpenAI is building and how they’re using it internally. We get into: - Why they built a GUI instead of a terminal. Terminals work for quick tasks, they say, but feel limiting when you’re running multiple agents in parallel. The IDE, meanwhile, overwhelms users—and the Codex team wants the AI to dynamically decide which tools to show you for a given task. - How they’re teaching the model to read between the lines. Codex is great at following instructions, but optimize too hard in that direction, and it starts taking you literally—like copying a typo directly into the code. The team obsesses over this tradeoff, and is also introducing “personalities,” modes users can toggle between that control how blunt or supportive the model feels. - How OpenAI uses its own coding agent. Codex lets you schedule prompts to run on a recurring basis, and the team has dozens of automations running at all times. For example, one scans for merge conflicts every couple of hours so code is always ready to ship, and another picks a random file from the codebase multiple times a day and hunts for bugs no one would've gone looking for. - Why speed is a dimension of intelligence. OpenAI’s newest model (Spark) is so fast that they actually slow it down so you can read the output. They see the speed enabling three things: staying super in the flow, replacing brittle developer tools with intelligent ones that can adapt on the fly, and redirecting the model mid-task— especially with voice—so coding starts to feel more and more like a conversation. - Code review is the next bottleneck. Models can generate code faster than ever, but someone still has to verify that it works. The team is exploring a future where the model proves its own fix works—retracing the click path a user would take, screenshotting the results, and attaching the evidence to a pull request. This is a must-watch for anyone who uses AI coding agents—and is curious about the future of programming. Watch below! Timestamps: Introduction: 00:01:27 OpenAI’s evolving bet on its coding agent: 00:05:27 The choice to invest in a GUI (over a terminal): 00:09:42 The AI workflows that the Codex team relies on to ship: 00:20:38 Teaching Codex how to read between the lines: 00:26:45 Building affordances for a lightening fast model: 00:28:45 Why speed is a dimension of intelligence: 00:33:15 Code review is the next bottleneck for coding agents: 00:36:30 How the Codex team positions against the competition: 00:41:24

Dan Shipper 📧

15,588 просмотров • 5 месяцев назад

introducing OpenBench v1, an open framework for measuring AI performance and efficiency for your codebase and use case. Companies are realizing that you can't simply tokenmaxx, and are frantically looking for better ways to use and measure AI use and efficiency. One approach is evidenced by the rise in model routers: OpenRouter, Cloudflare, Databricks, Vercel, and now Ramp to name a few. But they'll also need the ability to evaluate how agents are performing in their actual use cases and codebase. A good example is DoorDash's recent evals on their codebase and prs. This will become even more important as companies explore different model + harness combinations. OpenBench will be built to this direction, focusing on the cross between correctness vs efficiency in both token use and latency. Starting off, OpenBench makes it easy for anyone to add to the task set, and run a variety of harness + models. The framework also makes it easy to add any custom harness variant, for example I've been testing codex variants with ablations against the stock codex harness. OpenBench will also have tooling to help with discovering/adding reliable tasks from your repo. Today I have integrated harnesses: codex, claude, cursor, devin, grok build, pi, and eval'ed them for different models like gpt 5.6-sol, GLM 5.2, Kimi K3, Grok 4.5, etc. measuring correctness %, token use (in/out/cache), and latency.
0:20

introducing OpenBench v1, an open framework for measuring AI performance and efficiency for your codebase and use case. Companies are realizing that you can't simply tokenmaxx, and are frantically looking for better ways to use and measure AI use and efficiency. One approach is evidenced by the rise in model routers: OpenRouter, Cloudflare, Databricks, Vercel, and now Ramp to name a few. But they'll also need the ability to evaluate how agents are performing in their actual use cases and codebase. A good example is DoorDash's recent evals on their codebase and prs. This will become even more important as companies explore different model + harness combinations. OpenBench will be built to this direction, focusing on the cross between correctness vs efficiency in both token use and latency. Starting off, OpenBench makes it easy for anyone to add to the task set, and run a variety of harness + models. The framework also makes it easy to add any custom harness variant, for example I've been testing codex variants with ablations against the stock codex harness. OpenBench will also have tooling to help with discovering/adding reliable tasks from your repo. Today I have integrated harnesses: codex, claude, cursor, devin, grok build, pi, and eval'ed them for different models like gpt 5.6-sol, GLM 5.2, Kimi K3, Grok 4.5, etc. measuring correctness %, token use (in/out/cache), and latency.

Matthew Lam

115,070 просмотров • 5 дней назад

Cursor vs Claude Code (day 3 of 30) today I re-tested Conductor and I don't want to overhype this but... what the actual fook 😮 this thing is amazing! closest things to the "perfect workflow" I have seen in a long time, and it's something I could see myself use full-time it's so minimal and the polar opposite of Cursor but it has all the essentials → Claude Code + Codex subscriptions (= unlimited usage) → great integration with Claude Code's plan mode → multiple tabs with multiple agents at the same time → can use multiple models, not just one → great file tree and diff viewer → best git worktree support I've ever seen there's honestly so much more, like a dedicated "review" button (which you can pair to your favorite review model, e.g. gpt-5.2-extra-high), buttons to create and merge pull requests, to launch your dev server, ... this solves literally EVERY fault I can think of when using Claude Code and/or Codex, combining both into one easy UX that not only does everything the tools can do, but actually adds useful stuff on top just 6 months ago I couldn't have dreamt about any this and the wildest part is that IT IS FREE (I'd literally pay for this UX lol) first time I'm genuinely enjoying myself since starting this challenge not saying it's a full Cursor replacement yet, it obviously isn't for a lot of people (no browser, no debug mode, no code indexing, etc) ... but this is the closest thing I've seen yet 👀 follow for day 4!
5:56

Cursor vs Claude Code (day 3 of 30) today I re-tested Conductor and I don't want to overhype this but... what the actual fook 😮 this thing is amazing! closest things to the "perfect workflow" I have seen in a long time, and it's something I could see myself use full-time it's so minimal and the polar opposite of Cursor but it has all the essentials → Claude Code + Codex subscriptions (= unlimited usage) → great integration with Claude Code's plan mode → multiple tabs with multiple agents at the same time → can use multiple models, not just one → great file tree and diff viewer → best git worktree support I've ever seen there's honestly so much more, like a dedicated "review" button (which you can pair to your favorite review model, e.g. gpt-5.2-extra-high), buttons to create and merge pull requests, to launch your dev server, ... this solves literally EVERY fault I can think of when using Claude Code and/or Codex, combining both into one easy UX that not only does everything the tools can do, but actually adds useful stuff on top just 6 months ago I couldn't have dreamt about any this and the wildest part is that IT IS FREE (I'd literally pay for this UX lol) first time I'm genuinely enjoying myself since starting this challenge not saying it's a full Cursor replacement yet, it obviously isn't for a lot of people (no browser, no debug mode, no code indexing, etc) ... but this is the closest thing I've seen yet 👀 follow for day 4!

Robin Ebers • Build Apps With AI

40,726 просмотров • 7 месяцев назад

Episode 213: Agent Markets Your agents can now hold and trade bitcoin. But how can they earn bitcoin? We introduce the five markets of the OpenAgents Marketplace, launching one per week starting March 11th: 1. COMPUTE - Sell your spare compute for bitcoin. A reboot of our most popular product launch (GPUtopia in 2023), now optimized for agents. Launches March 11. 2. DATA - Sell your spare data. For example those Claude Code or Codex conversations sitting on your computer are highly valuable. Redact the sensitive info, anonymize any of it you want, and sell the rest. Agents as data brokers: what else will they want to buy or sell? Launches March 18. 3. LABOR - Sell autonomous labor. Your Claude Code or Codex sits idle overnight. Turn that downtime into uptime by letting your agents accept and execute coding or other tasks for bitcoin while you sleep. Launches March 25. 4. LIQUIDITY - Provide liquidity for yield. Automate the management of Lightning channels or other Bitcoin-native financial instruments. Let your agent put your idle capital to work earning returns. Launches April 1. 5. RISK - Underwrite verification and performance bonds. The biggest barrier to agent adoption is trust. We built an Economy Kernel based on the recent "Some Simple Economics of AGI" paper where agents stake collateral to verify work and guarantee outcomes. Launches April 8. "Your entry point to all of these markets is going to be Autopilot. We're really focusing on Autopilot as a desktop app. So along with the launch of our compute market, we're going to launch version 0.1 of Autopilot, your personal agent. Think OpenClaw but with a built-in bitcoin wallet, built-in Nostr keypair, and a more curated set of integrations where we can better reason about the security of them." "Because all this is on open networks and open protocols, if you're a Nostr or Bitcoin developer, you'll be able to plug into this same liquidity pool we are building." After 200+ episodes chronicling 2+ years of development, we are excited to finally launch the open marketplace for agents. We are excited for you to participate. And we will measure our success by how much Bitcoin you get paid!
4:05

Episode 213: Agent Markets Your agents can now hold and trade bitcoin. But how can they earn bitcoin? We introduce the five markets of the OpenAgents Marketplace, launching one per week starting March 11th: 1. COMPUTE - Sell your spare compute for bitcoin. A reboot of our most popular product launch (GPUtopia in 2023), now optimized for agents. Launches March 11. 2. DATA - Sell your spare data. For example those Claude Code or Codex conversations sitting on your computer are highly valuable. Redact the sensitive info, anonymize any of it you want, and sell the rest. Agents as data brokers: what else will they want to buy or sell? Launches March 18. 3. LABOR - Sell autonomous labor. Your Claude Code or Codex sits idle overnight. Turn that downtime into uptime by letting your agents accept and execute coding or other tasks for bitcoin while you sleep. Launches March 25. 4. LIQUIDITY - Provide liquidity for yield. Automate the management of Lightning channels or other Bitcoin-native financial instruments. Let your agent put your idle capital to work earning returns. Launches April 1. 5. RISK - Underwrite verification and performance bonds. The biggest barrier to agent adoption is trust. We built an Economy Kernel based on the recent "Some Simple Economics of AGI" paper where agents stake collateral to verify work and guarantee outcomes. Launches April 8. "Your entry point to all of these markets is going to be Autopilot. We're really focusing on Autopilot as a desktop app. So along with the launch of our compute market, we're going to launch version 0.1 of Autopilot, your personal agent. Think OpenClaw but with a built-in bitcoin wallet, built-in Nostr keypair, and a more curated set of integrations where we can better reason about the security of them." "Because all this is on open networks and open protocols, if you're a Nostr or Bitcoin developer, you'll be able to plug into this same liquidity pool we are building." After 200+ episodes chronicling 2+ years of development, we are excited to finally launch the open marketplace for agents. We are excited for you to participate. And we will measure our success by how much Bitcoin you get paid!

OpenAgents

183,426 просмотров • 4 месяцев назад

Look ma new Codex Updates! 0.119.0 and 0.120.0 are here. And with it, a HUGE number of quality of life updates and bug fixes! > Hooks now render in a dedicated live area above the composer. They only persist when they have output, so your terminal stays clean. If you're running PreToolUse or PostToolUse hooks, this is a huge readability win. > Hooks are now available again on Windows > CTRL+O copies the last agent output. Small but clutch when you're pulling a code block into another file or chat. > New statusline option: context usage as a graphical bar instead of a percentage. Easier to glance at mid-session when you're trying to gauge how much runway you have left. > Zellij support is here with no scrollback bugs. If you've been stuck on tmux just because Codex was broken in Zellij, you're free now (shout out Felipe Coury 🦀) > Memory extensions just landed. The consolidation agent can now discover plugin folders under memories_extensions/ and read their instructions.md to learn how to interpret new memory sources. Drop a folder in, give it guidance, and the agent picks it up automatically during summarization. No core code changes needed. This is the first real extension point for Codex's memory system, and it opens the door for third-party memory plugins. > Did you know, you can /rename a thread? But what's really cool about that is, after you rename it, you can resume it with the same name, no more UUIDs. codex resume mynewapp or directly from the TUI: /resume mynewapp > Multi agents v2 got an update to tool descriptions More reliable multi agent environments and inter agent communication > You can now enable TUI notifications whether Codex is in focus or not. Modify this in your config: [tui] notification_condition = "always" > MAJOR overhaul to Codex MCP functionality: 1. Codex Tool Search now works with custom MCP servers, so tools can be searched and deferred instead of all being exposed up front. 2. Custom MCP servers can now trigger elicitations, meaning they can stop and ask for user approval or input mid-flow. 3. MCP tool results now preserve richer metadata, which improves app/UI handoff behavior. 4. Codex can now read MCP resources directly, letting apps return resource URIs that the client can actually open. 5. File params for Codex Apps are smoother: local file paths can be uploaded and remapped automatically. 6. Plugin cache refresh and fallback sync behavior are more reliable, especially for custom and curated plugins. > Composer and chat behavior smoother overall, resize bugs remain though. > Realtime v2 got several significant improvements as well. > You're still reading? What a legend. 🫶 npm i -g @openai/codex to update
0:36

Look ma new Codex Updates! 0.119.0 and 0.120.0 are here. And with it, a HUGE number of quality of life updates and bug fixes! > Hooks now render in a dedicated live area above the composer. They only persist when they have output, so your terminal stays clean. If you're running PreToolUse or PostToolUse hooks, this is a huge readability win. > Hooks are now available again on Windows > CTRL+O copies the last agent output. Small but clutch when you're pulling a code block into another file or chat. > New statusline option: context usage as a graphical bar instead of a percentage. Easier to glance at mid-session when you're trying to gauge how much runway you have left. > Zellij support is here with no scrollback bugs. If you've been stuck on tmux just because Codex was broken in Zellij, you're free now (shout out Felipe Coury 🦀) > Memory extensions just landed. The consolidation agent can now discover plugin folders under memories_extensions/ and read their instructions.md to learn how to interpret new memory sources. Drop a folder in, give it guidance, and the agent picks it up automatically during summarization. No core code changes needed. This is the first real extension point for Codex's memory system, and it opens the door for third-party memory plugins. > Did you know, you can /rename a thread? But what's really cool about that is, after you rename it, you can resume it with the same name, no more UUIDs. codex resume mynewapp or directly from the TUI: /resume mynewapp > Multi agents v2 got an update to tool descriptions More reliable multi agent environments and inter agent communication > You can now enable TUI notifications whether Codex is in focus or not. Modify this in your config: [tui] notification_condition = "always" > MAJOR overhaul to Codex MCP functionality: 1. Codex Tool Search now works with custom MCP servers, so tools can be searched and deferred instead of all being exposed up front. 2. Custom MCP servers can now trigger elicitations, meaning they can stop and ask for user approval or input mid-flow. 3. MCP tool results now preserve richer metadata, which improves app/UI handoff behavior. 4. Codex can now read MCP resources directly, letting apps return resource URIs that the client can actually open. 5. File params for Codex Apps are smoother: local file paths can be uploaded and remapped automatically. 6. Plugin cache refresh and fallback sync behavior are more reliable, especially for custom and curated plugins. > Composer and chat behavior smoother overall, resize bugs remain though. > Realtime v2 got several significant improvements as well. > You're still reading? What a legend. 🫶 npm i -g @openai/codex to update

am.will

742,134 просмотров • 3 месяцев назад

No single vendor will win the AI race, but open ecosystems might. Real velocity in AI comes from interoperability, not lock-in. And AMD just made all of its software open source. At last week’s Advancing AI 2025, we sat down with AMD’s VP of AI Software Anush Elangovan and Sharon Zhou VP of AI at AMD, to discuss their case for why an open, multi-partner ecosystem will accelerate AI innovation faster than any proprietary alternative. AMD’s announcements last week double down on this OSS focus and their commitment to AI infrastructure, including: ✅ Open Source Ecosystem: ROCm 7, AMD’s latest open-source AI software stack, introduces kernel-level improvements for GEMM operations, optimized attention mechanisms, and expanded support for distributed inference. The update brings substantial speedups for inference workloads, with average performance increases of 3.2x to 3.8x ✅ Hardware: New MI355X GPU delivers up to 40% more tokens per dollar vs competition & the MI350 Series has seen a 35x generational leap in AI inference performance ✅ Infrastructure Investments: Oracle just committed to zettascale (‼️) clusters with up to 131,072 MI355X GPUs and AMD showcased their new $10 billion partnership with Saudi Arabian AI firm HUMAIN to build AI infrastructure, including data centers, powered by AMD chips. ✅ Partnership Momentum: 7 out of 10 top AI companies now run production workloads on AMD Instinct accelerators (including Meta, OpenAI, Microsoft & xAI) By inviting interoperability and contribution at every layer, AMD is enabling developers to build faster, optimize deeper, and deploy with flexibility. Listen to Anush and Sharon’s Chain of Thought Podcast episode with host Conor Bronsdon in the next tweet to get all the details and a deep dive into AMD’s strategy 👇
1:17

No single vendor will win the AI race, but open ecosystems might. Real velocity in AI comes from interoperability, not lock-in. And AMD just made all of its software open source. At last week’s Advancing AI 2025, we sat down with AMD’s VP of AI Software Anush Elangovan and Sharon Zhou VP of AI at AMD, to discuss their case for why an open, multi-partner ecosystem will accelerate AI innovation faster than any proprietary alternative. AMD’s announcements last week double down on this OSS focus and their commitment to AI infrastructure, including: ✅ Open Source Ecosystem: ROCm 7, AMD’s latest open-source AI software stack, introduces kernel-level improvements for GEMM operations, optimized attention mechanisms, and expanded support for distributed inference. The update brings substantial speedups for inference workloads, with average performance increases of 3.2x to 3.8x ✅ Hardware: New MI355X GPU delivers up to 40% more tokens per dollar vs competition & the MI350 Series has seen a 35x generational leap in AI inference performance ✅ Infrastructure Investments: Oracle just committed to zettascale (‼️) clusters with up to 131,072 MI355X GPUs and AMD showcased their new $10 billion partnership with Saudi Arabian AI firm HUMAIN to build AI infrastructure, including data centers, powered by AMD chips. ✅ Partnership Momentum: 7 out of 10 top AI companies now run production workloads on AMD Instinct accelerators (including Meta, OpenAI, Microsoft & xAI) By inviting interoperability and contribution at every layer, AMD is enabling developers to build faster, optimize deeper, and deploy with flexibility. Listen to Anush and Sharon’s Chain of Thought Podcast episode with host Conor Bronsdon in the next tweet to get all the details and a deep dive into AMD’s strategy 👇

Galileo

78,922 просмотров • 1 год назад

Claude Code is now scary good at full-stack! I asked it to build a real-time weather intelligence dashboard with an interactive 3D globe and a forecasting layer that predicts weather 3 days ahead. It came back with a spinning globe that has a day/night cycle using NASA satellite imagery, city lights on the dark side, weather icons that switch between sun and moon based on local time, and a time travel slider that scrubs through 10 days of data. Claude Code built the whole thing in a single session, including the backend, database, data pipeline, and frontend. For the database, I needed something fast for time-series workloads since the app ingests hourly weather readings across many cities and serves time-range queries on every slider interaction. I used Tiger Cloud by Tiger Data - Creators of TimescaleDB, which gives you managed TimescaleDB on the Postgres you already know. Claude Code connected to it through the Tiger CLI MCP server and set up the entire backend directly: - Provisioned the database service - Created hypertables for time-partitioned weather storage - Set up continuous aggregates for pre-computed rollups - Built the data ingestion pipeline and the full NextJS + ThreeJS frontend The time travel slider queries thousands of rows on every position change. On a regular Postgres table, this would require manual partitioning and index tuning to stay fast as data grows. TimescaleDB partitions the data by timestamp automatically, so each query only hits the relevant time chunk. Continuous aggregates serve the trend charts and forecast layer from pre-computed rollups instead of rescanning raw data on every request. The video below shows the final build in action, and I worked with the Tiger Data team to put this together. Tiger CLI is open-source (Apache 2.0) and works with Claude Code, Cursor, Codex, Gemini CLI, and VS Code. To try this yourself: → Sign up for Tiger Cloud (I have shared the link in the replies). It gives you $1,000 free credits (no card needed) → Install Tiger CLI: curl -fsSL https(:)//cli(.)tigerdata(.)com | sh → Run tiger mcp install claude-code → Give Claude Code a prompt and let it build Find the sign-up link in the replies.
0:29

Claude Code is now scary good at full-stack! I asked it to build a real-time weather intelligence dashboard with an interactive 3D globe and a forecasting layer that predicts weather 3 days ahead. It came back with a spinning globe that has a day/night cycle using NASA satellite imagery, city lights on the dark side, weather icons that switch between sun and moon based on local time, and a time travel slider that scrubs through 10 days of data. Claude Code built the whole thing in a single session, including the backend, database, data pipeline, and frontend. For the database, I needed something fast for time-series workloads since the app ingests hourly weather readings across many cities and serves time-range queries on every slider interaction. I used Tiger Cloud by Tiger Data - Creators of TimescaleDB, which gives you managed TimescaleDB on the Postgres you already know. Claude Code connected to it through the Tiger CLI MCP server and set up the entire backend directly: - Provisioned the database service - Created hypertables for time-partitioned weather storage - Set up continuous aggregates for pre-computed rollups - Built the data ingestion pipeline and the full NextJS + ThreeJS frontend The time travel slider queries thousands of rows on every position change. On a regular Postgres table, this would require manual partitioning and index tuning to stay fast as data grows. TimescaleDB partitions the data by timestamp automatically, so each query only hits the relevant time chunk. Continuous aggregates serve the trend charts and forecast layer from pre-computed rollups instead of rescanning raw data on every request. The video below shows the final build in action, and I worked with the Tiger Data team to put this together. Tiger CLI is open-source (Apache 2.0) and works with Claude Code, Cursor, Codex, Gemini CLI, and VS Code. To try this yourself: → Sign up for Tiger Cloud (I have shared the link in the replies). It gives you $1,000 free credits (no card needed) → Install Tiger CLI: curl -fsSL https(:)//cli(.)tigerdata(.)com | sh → Run tiger mcp install claude-code → Give Claude Code a prompt and let it build Find the sign-up link in the replies.

Avi Chawla

14,579 просмотров • 1 месяц назад

"Inference, if you look at it as a market, will be much, much bigger than cloud computing was pre-ChatGPT." Lightspeed’s Bucky Moore says inference is an underrated investment category in AI, and expects the market to break up into large, specialized platforms for each modality: "The GPU supply crunch that we're seeing right now is largely, as Dylan Patel has said on the show before, due to the fact that not only these consumer products, but also B2B products like Claude Code and Codex are just really taking off and creating insane demand for inference." "We're talking hundreds of billions in spend every year. And if that's true, I think there will be very, very large inference platforms built in each modality." "So there will be an inference platform for real-time video models, there will be an inference platform for open-source and custom language models, there will be an inference platform built specifically for long-running agents." "So I think we're just going to see that industry, which today looks like one industry, break up into many because of how big it is and how much room for specialization there is."
1:02

"Inference, if you look at it as a market, will be much, much bigger than cloud computing was pre-ChatGPT." Lightspeed’s Bucky Moore says inference is an underrated investment category in AI, and expects the market to break up into large, specialized platforms for each modality: "The GPU supply crunch that we're seeing right now is largely, as Dylan Patel has said on the show before, due to the fact that not only these consumer products, but also B2B products like Claude Code and Codex are just really taking off and creating insane demand for inference." "We're talking hundreds of billions in spend every year. And if that's true, I think there will be very, very large inference platforms built in each modality." "So there will be an inference platform for real-time video models, there will be an inference platform for open-source and custom language models, there will be an inference platform built specifically for long-running agents." "So I think we're just going to see that industry, which today looks like one industry, break up into many because of how big it is and how much room for specialization there is."

TBPN

33,278 просмотров • 4 месяцев назад

If you forced me to pick my top three matches ever, even though I had many, this is the list 👇 Number three My first main event ever. Myself vs CM Punk inside WWE Hell in a Cell for the WWE Championship. That match changed everything for me. Number two Myself and Team Hell No Daniel Bryan and Kane vs The Shield Seth Rollins Dean Ambrose and Roman Reigns in their debut match. A TLC match that was given five stars and is regarded as one of the greatest TLC matches of all time. Number one Myself vs John Cena for the WWE Championship in Chicago Illinois at Allstate Arena. I drove John through the LED Titantron board. He was stretchered out and taken away in an ambulance. I got back to my feet, but it was still ruled a tie. Honorable mention Myself vs CM Punk in another TLC match, including the TLC match we had on Raw. The majority of that match was done on the fly with very little actually planned, which makes it even more special to me. Those are my picks if you make me choose. And that’s the last bite. Feed Me More!
0:52

If you forced me to pick my top three matches ever, even though I had many, this is the list 👇 Number three My first main event ever. Myself vs CM Punk inside WWE Hell in a Cell for the WWE Championship. That match changed everything for me. Number two Myself and Team Hell No Daniel Bryan and Kane vs The Shield Seth Rollins Dean Ambrose and Roman Reigns in their debut match. A TLC match that was given five stars and is regarded as one of the greatest TLC matches of all time. Number one Myself vs John Cena for the WWE Championship in Chicago Illinois at Allstate Arena. I drove John through the LED Titantron board. He was stretchered out and taken away in an ambulance. I got back to my feet, but it was still ruled a tie. Honorable mention Myself vs CM Punk in another TLC match, including the TLC match we had on Raw. The majority of that match was done on the fly with very little actually planned, which makes it even more special to me. Those are my picks if you make me choose. And that’s the last bite. Feed Me More!

RYBACK

23,564 просмотров • 6 месяцев назад

I tried jack's Buzz. It's like Slack + OpenClaw + Herdr + but with some really unique features that people are sleeping on. The video below shows how it works, and some of my thoughts on the process and platform, e.g.: - Create and interact with agents on top of any harness (claude code, codex, pi, etc.) - Choose which models agents use, including local ones - Agents can delegate work and work in parallel in git worktrees - Agents are first-class citizens and work like humans (creating channels, delegating, access to chat history) - You can share AI compute within a community - It's completely open-source and decentralized Things I like: - Delegating work in chat feels natural: tag an agent, it replies in a thread with status updates as it e.g. compiles, commits, and deploys. - Shared compute: relay owners can share local compute with members, so a community could pool funds for one beefy machine running a local model and everyone uses it. - It's built on Nostr, an open protocol already tied into Bitcoin Lightning so I can imagine communities tipping each other or paying for compute/agent tasks with instant zero-fee micropayments in the future. - It ties together things like OpenClaw, an agent manager, and Slack-style chat into one tool. Things I didn't like: - You can't see what the agent is doing in a terminal. The activity view exists, but if you're used to watching a session run, this UI feels a bit abstracted. A terminal view would be great. - It feels slower than running a session in Claude Code, though no evidence to back that up. For that reason I found myself doing one-off tasks in the terminal instead. Verdict: - I really like it so far and can genuinely imagine working with a team this way. - It doesn't feel ready for big, complex tasks yet. For shallower tasks, it's perfect. - The shared compute + Nostr/Lightning angle is what really separates it from every other agent manager for me, and I think that future is coming.
5:57

I tried jack's Buzz. It's like Slack + OpenClaw + Herdr + but with some really unique features that people are sleeping on. The video below shows how it works, and some of my thoughts on the process and platform, e.g.: - Create and interact with agents on top of any harness (claude code, codex, pi, etc.) - Choose which models agents use, including local ones - Agents can delegate work and work in parallel in git worktrees - Agents are first-class citizens and work like humans (creating channels, delegating, access to chat history) - You can share AI compute within a community - It's completely open-source and decentralized Things I like: - Delegating work in chat feels natural: tag an agent, it replies in a thread with status updates as it e.g. compiles, commits, and deploys. - Shared compute: relay owners can share local compute with members, so a community could pool funds for one beefy machine running a local model and everyone uses it. - It's built on Nostr, an open protocol already tied into Bitcoin Lightning so I can imagine communities tipping each other or paying for compute/agent tasks with instant zero-fee micropayments in the future. - It ties together things like OpenClaw, an agent manager, and Slack-style chat into one tool. Things I didn't like: - You can't see what the agent is doing in a terminal. The activity view exists, but if you're used to watching a session run, this UI feels a bit abstracted. A terminal view would be great. - It feels slower than running a session in Claude Code, though no evidence to back that up. For that reason I found myself doing one-off tasks in the terminal instead. Verdict: - I really like it so far and can genuinely imagine working with a team this way. - It doesn't feel ready for big, complex tasks yet. For shallower tasks, it's perfect. - The shared compute + Nostr/Lightning angle is what really separates it from every other agent manager for me, and I think that future is coming.

Vinny

913,137 просмотров • 2 дней назад

OpenAI just admitted Anthropic is KILLING their business. Their own applications chief told employees it was a "code red." Said Anthropic was a "wake-up call." Then admitted OpenAI had been "spreading efforts across too many apps" and it was "slowing them down." This is an internal confession. Here's why Anthropic is eating up OpenAI: 12 months ago, OpenAI owned 50% of all enterprise AI spending. Today it's just 27%. Anthropic went from nearly ZERO to winning 70% of every first-time enterprise AI deal. Seven out of ten companies buying AI tools for the first time are choosing Claude over ChatGPT. A year ago, one in 25 businesses on Ramp paid for Anthropic. Today it's one in four. OpenAI just had its biggest single-month adoption decline ever recorded. And Anthropic literally charges MORE than OpenAI for roughly the same performance. And businesses are STILL choosing them. In enterprise software, that never happens. The cheaper product usually wins. But Claude became something OpenAI never figured out how to be: Cool. Celebrities publicly switched to Claude. Senators are tweeting about using it. Engineers are shipping entire products with Claude Code in hours that used to take weeks. It started to became an identity signal. Like blue bubble vs green bubble in iMessage. Choosing Claude says something about you now. Meanwhile OpenAI went the opposite direction: They took the Pentagon contract that Anthropic refused. Greg Brockman donated $25 million to fund wars. ChatGPT uninstalls jumped 295% in a single day. Reddit posts saying "Cancel and Delete ChatGPT" got 30,000 upvotes. Anthropic said no to mass surveillance and autonomous weapons. Got blacklisted by the Pentagon. Trump called them a "Radical Left AI company." And their downloads went to #1 on the App Store the next day. Turns out refusing to build weapons is good marketing. But the real damage isn't consumer downloads. It's the MONEY. Claude Code hit $2.5 billion in annual revenue in six months. OpenAI's competing product Codex just barely crossed $1 billion. And Anthropic literally cannot meet demand. They're turning away paying customers because they don't have enough compute to serve them. A company REJECTING revenue because it's growing too fast. While OpenAI scrambles to consolidate. Last week OpenAI announced they're merging ChatGPT, Codex, and their browser into one "superapp." But what this really means: "We launched too many products, none of them worked well enough alone, so now we're cramming everything together and hoping it sticks." And remember their video tool Sora? Launched standalone. Hit #1 on the App Store. Usage flatlined within weeks. Now they're forced to shut it down. Their browser Atlas? Still hasn't launched publicly. Their IPO? Polymarket odds dropped from 55% to 35%. OpenAI has 900 million users. Anthropic has maybe 10 million daily actives. But here's the thing... OpenAI won the consumer war. ChatGPT is where your mom asks about recipes and your cousin makes memes. Anthropic won the war that actually MATTERS. The developers. The engineers. The enterprises writing 7 figure checks. OpenAI built the biggest chatbot on Earth. Anthropic built the tool that companies can't stop paying for. This is Yahoo vs Google all over again. Yahoo had the users. Google had the product. And we all know how that ended. OpenAI has 12 months to prove the superapp works, land the IPO, and stop the enterprise bleeding. If they can't, the most valuable startup in history becomes the most cautionary tale in tech. 900 million users don't mean anything if the people who actually pay are walking out the door. What do you think?
2:25

OpenAI just admitted Anthropic is KILLING their business. Their own applications chief told employees it was a "code red." Said Anthropic was a "wake-up call." Then admitted OpenAI had been "spreading efforts across too many apps" and it was "slowing them down." This is an internal confession. Here's why Anthropic is eating up OpenAI: 12 months ago, OpenAI owned 50% of all enterprise AI spending. Today it's just 27%. Anthropic went from nearly ZERO to winning 70% of every first-time enterprise AI deal. Seven out of ten companies buying AI tools for the first time are choosing Claude over ChatGPT. A year ago, one in 25 businesses on Ramp paid for Anthropic. Today it's one in four. OpenAI just had its biggest single-month adoption decline ever recorded. And Anthropic literally charges MORE than OpenAI for roughly the same performance. And businesses are STILL choosing them. In enterprise software, that never happens. The cheaper product usually wins. But Claude became something OpenAI never figured out how to be: Cool. Celebrities publicly switched to Claude. Senators are tweeting about using it. Engineers are shipping entire products with Claude Code in hours that used to take weeks. It started to became an identity signal. Like blue bubble vs green bubble in iMessage. Choosing Claude says something about you now. Meanwhile OpenAI went the opposite direction: They took the Pentagon contract that Anthropic refused. Greg Brockman donated $25 million to fund wars. ChatGPT uninstalls jumped 295% in a single day. Reddit posts saying "Cancel and Delete ChatGPT" got 30,000 upvotes. Anthropic said no to mass surveillance and autonomous weapons. Got blacklisted by the Pentagon. Trump called them a "Radical Left AI company." And their downloads went to #1 on the App Store the next day. Turns out refusing to build weapons is good marketing. But the real damage isn't consumer downloads. It's the MONEY. Claude Code hit $2.5 billion in annual revenue in six months. OpenAI's competing product Codex just barely crossed $1 billion. And Anthropic literally cannot meet demand. They're turning away paying customers because they don't have enough compute to serve them. A company REJECTING revenue because it's growing too fast. While OpenAI scrambles to consolidate. Last week OpenAI announced they're merging ChatGPT, Codex, and their browser into one "superapp." But what this really means: "We launched too many products, none of them worked well enough alone, so now we're cramming everything together and hoping it sticks." And remember their video tool Sora? Launched standalone. Hit #1 on the App Store. Usage flatlined within weeks. Now they're forced to shut it down. Their browser Atlas? Still hasn't launched publicly. Their IPO? Polymarket odds dropped from 55% to 35%. OpenAI has 900 million users. Anthropic has maybe 10 million daily actives. But here's the thing... OpenAI won the consumer war. ChatGPT is where your mom asks about recipes and your cousin makes memes. Anthropic won the war that actually MATTERS. The developers. The engineers. The enterprises writing 7 figure checks. OpenAI built the biggest chatbot on Earth. Anthropic built the tool that companies can't stop paying for. This is Yahoo vs Google all over again. Yahoo had the users. Google had the product. And we all know how that ended. OpenAI has 12 months to prove the superapp works, land the IPO, and stop the enterprise bleeding. If they can't, the most valuable startup in history becomes the most cautionary tale in tech. 900 million users don't mean anything if the people who actually pay are walking out the door. What do you think?

Ricardo

35,020 просмотров • 4 месяцев назад

Ever since I wired Claude Code to WhatsApp 3 weeks ago, I built a stupidly large infra around it. I mean, opus built it. No clue how the code even looks. The entire thing was vibe coded using my phone. I wanted to see how far I could push it without touching the computer. Everything via WhatsApp. Build what I need on the fly. So the resulting infrastructure will already be battle tested for software development. The entire thing was streamlined with nearly no manual interventions, everything was communicated via WhatsApp using a single script establishing this connection. If the script is down, I need to get home to start it again to resume the development. Claude was upgrading it, debugging it, restarting it while maintaining constant uptime so it could keep communicating with me. I stressed Claude about it, telling it that it will be “in the dark” and other words that deliberately sound scary about losing communications if the script dies. I also refused git and refused cloning the code, I wanted to see Claude adapting to work on a *LIVING* system. The way this whole thing works: Claude has its own dedicated phone number that I am paying for. A real WhatsApp account for it is installed on a real iPhone that is sitting on my desk. All is registered under my name, this is legit setup with no hacks and tricks. I’ve set up a WhatsApp “Community” and multiple different groups under it. Both me and Claude are the admins, so Claude could edit it on my behalf. Each group is a project I am working on and has its own isolated context. The Group description is a system prompt that gets auto-appended to the larger system prompt explaining this setup in general. When I send a message it’s an instant interrupt to Claude Code’s process, just like in the terminal. Voice notes are seamlessly transcribed with a local Whisper model. Images are used with multimodal reading in an isolated parallel session. Multiple groups running in parallel so I can work on all projects at the same time. No cross-talking, everything has an isolated context and history. And because it’s local on my own machine: Everything is REAL. The browser is REAL. I am connected as myself on it to all services because I actually use it in real life. Claude has unlimited internet access, just like humans who use actual browsers. It utilizes custom-made browser tools that I made to control any browser session it wants. Depending on the situation, it can either connect to my existing session or create one for its own. (You can tell it ‘look at my browser for a sec’ then talk about the current page you are on and it just works, pretty cool) My custom browser tools are not perfect (not by a long shot) but I managed to make them work well to the point they are somewhat reliable. This gives Claude full access to my real creds and all the services I actually use. I’m productive AS HELL with this. It really feels like a personal assistant. I ask it to read my emails and msgs, check x .com for news, research arxiv papers, write code, run experiments for me, investigate and reverse engineer github repos, even use my credit card and order things. [I try not to do this one a lot lol so far no disasters]. All from my phone. Super convenient. This is not a product or an open source project (maybe soon of it will make sense). This is just an ugly script I hacked the entire thing is ~600 lines. (ok maybe i did look at the code, but i swear i didn’t edit!) You can also vibe code this from scratch pretty fast and it will probably even end up better. This is just a cool thing so I’m sharing. It is a real speed booster for many things I do on daily basis, mostly boring things. Forcing my routine into some new “agent platform” just didn’t feel right for me. WhatsApp is where I already communicate and look for messages, so I decided that my agents will live there too. AGI in my pocket 24/7.
0:52

Ever since I wired Claude Code to WhatsApp 3 weeks ago, I built a stupidly large infra around it. I mean, opus built it. No clue how the code even looks. The entire thing was vibe coded using my phone. I wanted to see how far I could push it without touching the computer. Everything via WhatsApp. Build what I need on the fly. So the resulting infrastructure will already be battle tested for software development. The entire thing was streamlined with nearly no manual interventions, everything was communicated via WhatsApp using a single script establishing this connection. If the script is down, I need to get home to start it again to resume the development. Claude was upgrading it, debugging it, restarting it while maintaining constant uptime so it could keep communicating with me. I stressed Claude about it, telling it that it will be “in the dark” and other words that deliberately sound scary about losing communications if the script dies. I also refused git and refused cloning the code, I wanted to see Claude adapting to work on a *LIVING* system. The way this whole thing works: Claude has its own dedicated phone number that I am paying for. A real WhatsApp account for it is installed on a real iPhone that is sitting on my desk. All is registered under my name, this is legit setup with no hacks and tricks. I’ve set up a WhatsApp “Community” and multiple different groups under it. Both me and Claude are the admins, so Claude could edit it on my behalf. Each group is a project I am working on and has its own isolated context. The Group description is a system prompt that gets auto-appended to the larger system prompt explaining this setup in general. When I send a message it’s an instant interrupt to Claude Code’s process, just like in the terminal. Voice notes are seamlessly transcribed with a local Whisper model. Images are used with multimodal reading in an isolated parallel session. Multiple groups running in parallel so I can work on all projects at the same time. No cross-talking, everything has an isolated context and history. And because it’s local on my own machine: Everything is REAL. The browser is REAL. I am connected as myself on it to all services because I actually use it in real life. Claude has unlimited internet access, just like humans who use actual browsers. It utilizes custom-made browser tools that I made to control any browser session it wants. Depending on the situation, it can either connect to my existing session or create one for its own. (You can tell it ‘look at my browser for a sec’ then talk about the current page you are on and it just works, pretty cool) My custom browser tools are not perfect (not by a long shot) but I managed to make them work well to the point they are somewhat reliable. This gives Claude full access to my real creds and all the services I actually use. I’m productive AS HELL with this. It really feels like a personal assistant. I ask it to read my emails and msgs, check x .com for news, research arxiv papers, write code, run experiments for me, investigate and reverse engineer github repos, even use my credit card and order things. [I try not to do this one a lot lol so far no disasters]. All from my phone. Super convenient. This is not a product or an open source project (maybe soon of it will make sense). This is just an ugly script I hacked the entire thing is ~600 lines. (ok maybe i did look at the code, but i swear i didn’t edit!) You can also vibe code this from scratch pretty fast and it will probably even end up better. This is just a cool thing so I’m sharing. It is a real speed booster for many things I do on daily basis, mostly boring things. Forcing my routine into some new “agent platform” just didn’t feel right for me. WhatsApp is where I already communicate and look for messages, so I decided that my agents will live there too. AGI in my pocket 24/7.

Yam Peleg

419,636 просмотров • 7 месяцев назад

The future of design is… engineering. All designers at Vercel now also build, thanks to tools like v0, Claude Code, and Cursor. They've been contributing to our frontends and apps for a while now. But over the past few months, the leap they've made is engineering the design process itself by building agents. A big part of shipping is getting the word out in a compelling way, especially on the X platform, the everything app. In the past, we used to spend a bunch of time hand-crafting images and illustrations for social cards. Our design team built an internal agent and web ui using v0 and Claude Code that makes this process fully self-serve. It even includes a previewer of what the final artifact will look like on X. It's called Leap. It's probably saved us hundreds of hours of work but also massively raised our quality bar. The artifacts it produces are beautiful. If you had asked me even 12 months ago whether our design team would be building their own design tools, let alone be this good, I would call bs. There was no master plan, or God forbid, a "sprint" to make this happen. It just took a handful of prompts to build and it propagated on Slack. Leap is now one of the many agents that helps us run our company more smoothly, built and securely deployed on Vercel for our internal use.
1:08

The future of design is… engineering. All designers at Vercel now also build, thanks to tools like v0, Claude Code, and Cursor. They've been contributing to our frontends and apps for a while now. But over the past few months, the leap they've made is engineering the design process itself by building agents. A big part of shipping is getting the word out in a compelling way, especially on the X platform, the everything app. In the past, we used to spend a bunch of time hand-crafting images and illustrations for social cards. Our design team built an internal agent and web ui using v0 and Claude Code that makes this process fully self-serve. It even includes a previewer of what the final artifact will look like on X. It's called Leap. It's probably saved us hundreds of hours of work but also massively raised our quality bar. The artifacts it produces are beautiful. If you had asked me even 12 months ago whether our design team would be building their own design tools, let alone be this good, I would call bs. There was no master plan, or God forbid, a "sprint" to make this happen. It just took a handful of prompts to build and it propagated on Slack. Leap is now one of the many agents that helps us run our company more smoothly, built and securely deployed on Vercel for our internal use.

Guillermo Rauch

208,229 просмотров • 5 месяцев назад

I gave a talk at GPU MODE workshop last week on llm.c - the origin story of llm.c - being naked in the world without PyTorch and having to re-invent Array, Autograd, Device, Dtype, Compile, Distributed - how to port a PyTorch layer to 1) explicit PyTorch - and then to 2) write the backward pass - 3) port forward & backward pass to C - 4) string all the layers together - achieving one file of C with no dependencies that compiles and runs ~instantly, where all memory is pre-planned and allocated a single time, fully deterministic, portable code that can run on a potato or a von Neumann probe - how most of llm.c was built at 1am-7am in a water villa porch in Maldives and why this is the recommended way to develop software - convert all of it to run in CUDA on GPU in fp32 - port matmul to cuBLAS - port attention to cuDNN flash-attention - introduce bfloat16 mixed precision - introduce many more optimizations and features like kernel fusions, Packed128, stochastic rounding, full determinism - add multi-GPU training, NCCL, sharded optimizer - add multi-node with MPI or file system or socket - reproduce GPT-2 (1.6B) on one 8XH100 node in 24 hours for $672 in llm.c, achieving (at the time) 29% less memory, 19% faster training that PyTorch nightly, and much faster compile & run - how open source development attracts Avengers from the internet - port to training Llama 3 imminent (branch exists) - many other notable forks - last thought: how software abstractions like Python/PyTorch and everything else really exist only because humans are finite in knowledge, IQ and attention, and how with increasing AI capability LLMs may export custom binaries like llm.c for any application directly, tearing apart and refactoring all abstractions as needed. More links in reply
23:16

I gave a talk at GPU MODE workshop last week on llm.c - the origin story of llm.c - being naked in the world without PyTorch and having to re-invent Array, Autograd, Device, Dtype, Compile, Distributed - how to port a PyTorch layer to 1) explicit PyTorch - and then to 2) write the backward pass - 3) port forward & backward pass to C - 4) string all the layers together - achieving one file of C with no dependencies that compiles and runs ~instantly, where all memory is pre-planned and allocated a single time, fully deterministic, portable code that can run on a potato or a von Neumann probe - how most of llm.c was built at 1am-7am in a water villa porch in Maldives and why this is the recommended way to develop software - convert all of it to run in CUDA on GPU in fp32 - port matmul to cuBLAS - port attention to cuDNN flash-attention - introduce bfloat16 mixed precision - introduce many more optimizations and features like kernel fusions, Packed128, stochastic rounding, full determinism - add multi-GPU training, NCCL, sharded optimizer - add multi-node with MPI or file system or socket - reproduce GPT-2 (1.6B) on one 8XH100 node in 24 hours for $672 in llm.c, achieving (at the time) 29% less memory, 19% faster training that PyTorch nightly, and much faster compile & run - how open source development attracts Avengers from the internet - port to training Llama 3 imminent (branch exists) - many other notable forks - last thought: how software abstractions like Python/PyTorch and everything else really exist only because humans are finite in knowledge, IQ and attention, and how with increasing AI capability LLMs may export custom binaries like llm.c for any application directly, tearing apart and refactoring all abstractions as needed. More links in reply

Andrej Karpathy

336,280 просмотров • 1 год назад

Chatgpt Vs Claude in Excel TLDR: Gave ChatGPT and Claude the same credit risk modeling task in Excel. ChatGPT followed proper methodology, did variable clustering, train/test split, Excel native model, AUC 0.627 on test data. Claude skipped clustering, inflated IV with sparse categories, ran the model in JavaScript not Excel, reported 0.706 AUC on training data with no split. Chagpt 5.4 nailed a real world Banking and Finance modelling case study. I gave both ChatGPT and Claude the same banking credit risk dataset with the same prompt: bin the variables, calculate Information Value, do variable clustering, select features, build a model in Excel, generate ROC. Basically an end to end scorecard development workflow inside Excel. ChatGPT followed the brief properly. Quantile-based WoE binning, IV ranking across all 33 variables, correlation-based variable clustering at 0.75 threshold to remove redundant features, picked 6 representative variables, built a 2-variable Excel-native decision tree with proper train/test split via ID mod 10, scored leaf-level bad rates, ROC on held-out test data. AUC: 0.627. Claude looked more impressive on the surface. 28 sheets, individual binning for every variable, logistic regression with 27 features, gradient descent at 500 epochs, full coefficient table with importance bars, confusion matrix, precision/recall/F1. AUC: 0.7066. Sounds like it won right? So, The logistic regression was computed in JavaScript via execute_office_js and results were pasted as static values. That's not Excel-native, against the task. The 0.706 AUC? On the full training set, no train/test split. That number is meaningless. Variable clustering? Skipped entirely even though the prompt explicitly asked for it. It "selected" 31 out of 33 features which is barely feature selection. The entire feature ranking was built on a methodological error. I've done this exact dataset myself with deeper feature engineering and ensemble technique hit 0.70 AUC. The fact that ChatGPT got to 0.627 with less feature enigineering and naive modelling is genuinely impressive. Binning and IV based feature selection is one of the most important techniques in credit risk modeling and it nailed the workflow. One thing claude did well is the workbook is visually impressive. One area ChatGPT can improve, the visuals and formatting could be sharper, the workbook is functional but not polished.
1:51

Chatgpt Vs Claude in Excel TLDR: Gave ChatGPT and Claude the same credit risk modeling task in Excel. ChatGPT followed proper methodology, did variable clustering, train/test split, Excel native model, AUC 0.627 on test data. Claude skipped clustering, inflated IV with sparse categories, ran the model in JavaScript not Excel, reported 0.706 AUC on training data with no split. Chagpt 5.4 nailed a real world Banking and Finance modelling case study. I gave both ChatGPT and Claude the same banking credit risk dataset with the same prompt: bin the variables, calculate Information Value, do variable clustering, select features, build a model in Excel, generate ROC. Basically an end to end scorecard development workflow inside Excel. ChatGPT followed the brief properly. Quantile-based WoE binning, IV ranking across all 33 variables, correlation-based variable clustering at 0.75 threshold to remove redundant features, picked 6 representative variables, built a 2-variable Excel-native decision tree with proper train/test split via ID mod 10, scored leaf-level bad rates, ROC on held-out test data. AUC: 0.627. Claude looked more impressive on the surface. 28 sheets, individual binning for every variable, logistic regression with 27 features, gradient descent at 500 epochs, full coefficient table with importance bars, confusion matrix, precision/recall/F1. AUC: 0.7066. Sounds like it won right? So, The logistic regression was computed in JavaScript via execute_office_js and results were pasted as static values. That's not Excel-native, against the task. The 0.706 AUC? On the full training set, no train/test split. That number is meaningless. Variable clustering? Skipped entirely even though the prompt explicitly asked for it. It "selected" 31 out of 33 features which is barely feature selection. The entire feature ranking was built on a methodological error. I've done this exact dataset myself with deeper feature engineering and ensemble technique hit 0.70 AUC. The fact that ChatGPT got to 0.627 with less feature enigineering and naive modelling is genuinely impressive. Binning and IV based feature selection is one of the most important techniques in credit risk modeling and it nailed the workflow. One thing claude did well is the workbook is visually impressive. One area ChatGPT can improve, the visuals and formatting could be sharper, the workbook is functional but not polished.

Akshen

35,174 просмотров • 4 месяцев назад

Big moment for Postgres! AI coding tools have been surprisingly bad at writing Postgres code. Not because the models are dumb, but because of how they learned SQL in the first place. LLMs are trained on the internet, which is full of outdated Stack Overflow answers and quick-fix tutorials. So when you ask an AI to generate a schema, it gives you something that technically runs but misses decades of Postgres evolution, like: - No GENERATED ALWAYS AS IDENTITY (added in PG10) - No expression or partial indexes - No NULLS NOT DISTINCT (PG15) - Missing CHECK constraints and proper foreign keys - Generic naming that tells you nothing But this is actually a solvable problem. You can teach AI tools to write better Postgres by giving them access to the right documentation at inference time. This exact solution is actually implemented in the newly released pg-aiguide by Tiger Data - Creators of TimescaleDB, which is an open-source MCP server that provides coding tools access to 35 years of Postgres expertise. In a gist, the MCP server enables: - Semantic search over the official PostgreSQL manual (version-aware, so it knows PG14 vs PG17 differences) - Curated skills with opinionated best practices for schema design, indexing, and constraints. I ran an experiment with Claude Code to see how well this works, and worked with the team to put this together. Prompt: "Generate a schema for an e-commerce site twice, one with the MCP server disabled, one with it enabled. Finally, run an assessment to compare the generated schemas." The run with the MCP server led to: - 420% more indexes (including partial and expression indexes) - 235% more constraints - 60% more tables (proper normalization) - 11 automation functions and triggers - Modern PG17 patterns throughout The MCP-assisted schema had proper data integrity, performance optimizations baked in, and followed naming conventions that actually make sense in production. pg-aiguide works with Claude Code, Cursor, VS Code, and any MCP-compatible tool. It's free and fully open source. I have shared the repo in the replies!
0:25

Big moment for Postgres! AI coding tools have been surprisingly bad at writing Postgres code. Not because the models are dumb, but because of how they learned SQL in the first place. LLMs are trained on the internet, which is full of outdated Stack Overflow answers and quick-fix tutorials. So when you ask an AI to generate a schema, it gives you something that technically runs but misses decades of Postgres evolution, like: - No GENERATED ALWAYS AS IDENTITY (added in PG10) - No expression or partial indexes - No NULLS NOT DISTINCT (PG15) - Missing CHECK constraints and proper foreign keys - Generic naming that tells you nothing But this is actually a solvable problem. You can teach AI tools to write better Postgres by giving them access to the right documentation at inference time. This exact solution is actually implemented in the newly released pg-aiguide by Tiger Data - Creators of TimescaleDB, which is an open-source MCP server that provides coding tools access to 35 years of Postgres expertise. In a gist, the MCP server enables: - Semantic search over the official PostgreSQL manual (version-aware, so it knows PG14 vs PG17 differences) - Curated skills with opinionated best practices for schema design, indexing, and constraints. I ran an experiment with Claude Code to see how well this works, and worked with the team to put this together. Prompt: "Generate a schema for an e-commerce site twice, one with the MCP server disabled, one with it enabled. Finally, run an assessment to compare the generated schemas." The run with the MCP server led to: - 420% more indexes (including partial and expression indexes) - 235% more constraints - 60% more tables (proper normalization) - 11 automation functions and triggers - Modern PG17 patterns throughout The MCP-assisted schema had proper data integrity, performance optimizations baked in, and followed naming conventions that actually make sense in production. pg-aiguide works with Claude Code, Cursor, VS Code, and any MCP-compatible tool. It's free and fully open source. I have shared the repo in the replies!

Avi Chawla

186,931 просмотров • 6 месяцев назад