Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Matt Maher tested frontier models in Cursor v. other harnesses. Cursor boosted model performance by 11% on average: Gemini: 52% → 57% GPT-5.4: 82% → 88% Opus: 77% → 93% His benchmark measures how well models implement a 100-feature PRD. Cursor consistently outperformed.

edwin

30,635 subscribers

912,650 views • 4 months ago •via X (Twitter)

Science & Technology

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

Elon Musk just made one if the biggest moves in taking over the programming industry “SpaceX just bought Cursor for $60 billion. Do you realize how big this is? SpaceX went public — the biggest IPO in history. $75 billion raised, almost a $2 trillion valuation and the first thing to do with that money? Buy the most popular AI coding tool on the planet. Here's why that changes everything. Elon now owns 3 layers: the compute, Colossus data centers, the models, Grok through xAI, and now the tool that developers actually use every day. It's the full stack. And here's what makes Cursor different from Claude Code or Codex. Cursor is model agnostic. You can run Claude in it, GPT, Gemini, whatever model you want. It's not locked to any one company, and now it has SpaceX's resources behind it. Cursor said they were bottlenecked by compute. Well, that bottleneck has just been removed. $4 billion in annual revenue, over half the Fortune 500 already uses it, and now it's backed by a $2 trillion company. OpenAI has Codex, Anthropic has Claude Code, and now Elon has Cursor.” Let me break this down in simple terms Elon Musk now controls more of the full AI picture: - Massive computers, power (data centers like Colossus) - Smart AI models (Grok from xAI) - The actual tool millions of developers use every day (Cursor) For every day users this means Faster and smarter apps and websites in the future. More developers using powerful AI tools means new apps, games, websites, and features get built quicker and cheaper. This means better video games, smoother streaming, smarter phone apps and better programs For Developers they can describe what they want in plain English (“make a feature that does X”) and the AI handles more of the heavy lifting

Elon Musk just made one if the biggest moves in taking over the programming industry “SpaceX just bought Cursor for $60 billion. Do you realize how big this is? SpaceX went public — the biggest IPO in history. $75 billion raised, almost a $2 trillion valuation and the first thing to do with that money? Buy the most popular AI coding tool on the planet. Here's why that changes everything. Elon now owns 3 layers: the compute, Colossus data centers, the models, Grok through xAI, and now the tool that developers actually use every day. It's the full stack. And here's what makes Cursor different from Claude Code or Codex. Cursor is model agnostic. You can run Claude in it, GPT, Gemini, whatever model you want. It's not locked to any one company, and now it has SpaceX's resources behind it. Cursor said they were bottlenecked by compute. Well, that bottleneck has just been removed. $4 billion in annual revenue, over half the Fortune 500 already uses it, and now it's backed by a $2 trillion company. OpenAI has Codex, Anthropic has Claude Code, and now Elon has Cursor.” Let me break this down in simple terms Elon Musk now controls more of the full AI picture: - Massive computers, power (data centers like Colossus) - Smart AI models (Grok from xAI) - The actual tool millions of developers use every day (Cursor) For every day users this means Faster and smarter apps and websites in the future. More developers using powerful AI tools means new apps, games, websites, and features get built quicker and cheaper. This means better video games, smoother streaming, smarter phone apps and better programs For Developers they can describe what they want in plain English (“make a feature that does X”) and the AI handles more of the heavy lifting

Wall Street Apes

213,267 views • 1 month ago

Battle of the new coding models: Cursor vs Cognition They both make big claims being near the frontier - how well do they pass the Golden Gate Bridge test? I've included 5x non-cherry picked generations from each, followed by examples from pre-release test of Gemini 3 Pro, GPT-5 and Sonnet 4.5. Generations are mostly one shot, but each had errors twice, which it then fixed in the next generation. My impressions: - Both are clearly way behind the frontier - Between the two, Windsurf's SWE-1.5 is a winner for me - no huge errors like two parts of the bridge are the wrong way round - But SWE-1.5 generations were quite low-fi and Cursor's composer did have some better generations - Overall GPT-5 seems like the best model around on these tests, and Gemini 3 Pro seems to have a lot of potential Music via Suno

Battle of the new coding models: Cursor vs Cognition They both make big claims being near the frontier - how well do they pass the Golden Gate Bridge test? I've included 5x non-cherry picked generations from each, followed by examples from pre-release test of Gemini 3 Pro, GPT-5 and Sonnet 4.5. Generations are mostly one shot, but each had errors twice, which it then fixed in the next generation. My impressions: - Both are clearly way behind the frontier - Between the two, Windsurf's SWE-1.5 is a winner for me - no huge errors like two parts of the bridge are the wrong way round - But SWE-1.5 generations were quite low-fi and Cursor's composer did have some better generations - Overall GPT-5 seems like the best model around on these tests, and Gemini 3 Pro seems to have a lot of potential Music via Suno

Peter Gostev

23,583 views • 9 months ago

🧑‍🚀 Day 5 of the Cursor #vibejam Proudly sponsored by Cursor + bolt.new + GLIF Today a little bit about Cursor 3, it's a completely rebuilt interface designed around agents, you can run multiple AI agents in parallel, hand off between local and cloud, and go from commit to merged PR. It's powered by Composer 2, their own frontier coding model that tops benchmarks at $0.50/M input tokens — and the Fast variant runs at 200 tokens/sec, more than 2x faster than other frontier models. Perfect for vibe coding your game jam entry. And edwin from Cursor is here to help you if you get stuck! Lots of activity again today again on here, the best games (or well previews) I saw today: 🍣 Sushi Belt by steeno 🌊 Surfing Simulator by Erik (this one is a basic demo but very promising as it has realistic water physics) 🦆 Grand Quack Auto by kyzo (or as I like to call it Bali Simulator) 🏛️ Roman Combat Sim by terry trusner (continued from yesterday because it's had a lot of progress I think) YOU HAVE 24 DAYS LEFT! Reply in this thread with updates on your current games to share your progress, and add tag #vibejam so I see and can include you in the daily tweet There's $35,000 in prizes for you to win, see threads below for more info. The Gold prize is $20,000, bronze is $10,000 and silver is $5,000! Wanna to participate? You can still start today and submit your game any time before May 1!

🧑‍🚀 Day 5 of the Cursor #vibejam Proudly sponsored by Cursor + bolt.new + GLIF Today a little bit about Cursor 3, it's a completely rebuilt interface designed around agents, you can run multiple AI agents in parallel, hand off between local and cloud, and go from commit to merged PR. It's powered by Composer 2, their own frontier coding model that tops benchmarks at $0.50/M input tokens — and the Fast variant runs at 200 tokens/sec, more than 2x faster than other frontier models. Perfect for vibe coding your game jam entry. And edwin from Cursor is here to help you if you get stuck! Lots of activity again today again on here, the best games (or well previews) I saw today: 🍣 Sushi Belt by steeno 🌊 Surfing Simulator by Erik (this one is a basic demo but very promising as it has realistic water physics) 🦆 Grand Quack Auto by kyzo (or as I like to call it Bali Simulator) 🏛️ Roman Combat Sim by terry trusner (continued from yesterday because it's had a lot of progress I think) YOU HAVE 24 DAYS LEFT! Reply in this thread with updates on your current games to share your progress, and add tag #vibejam so I see and can include you in the daily tweet There's $35,000 in prizes for you to win, see threads below for more info. The Gold prize is $20,000, bronze is $10,000 and silver is $5,000! Wanna to participate? You can still start today and submit your game any time before May 1!

@levelsio

79,700 views • 3 months ago

Today's Training Data episode takes us BTS on the infrastructure challenges required to do large RL runs at scale, featuring Federico Cassano (Composer Lead at Cursor) and Dmytro Dzhulgakov (Co-Founder at Fireworks AI). The Cursor team trained Composer 2 on Fireworks by starting with a strong base model (Kimi 2.5) and performing large-scale mid-training on code tokens and web data to learn common patterns and libraries, followed by a large-scale Reinforcement Learning run to learn how to navigate the Cursor harness, call tools, and write correct code. Today's episode dives into the systems and infrastructure challenges of making that large RL run happening, and there were many (!!), from numerical mismatch to global distribution to synchronizing rollouts across asynchronous pipelines to keeping track of expert activation across runs and more. Extremely nerdy in-the-weeds challenges that Federico and Dima were delighted to nerd out on together :) Beyond RL infra, we also discussed Online vs Simulated rollouts, self-summarization for long-horizon agents, environment design ("the most powerful RL environment is the product itself"), and other technical nuggets. PS: We filmed this episode before the SpaceX news, while the Cursor team was still compute-constrained. While Cursor now has *all* the flops, the takeaways and hurdles crossed ring true for any serious application-level company that is racing to post-train their own models. I believe that more serious application companies will go the way of Cursor and post-train their own models. 00:00 Introduction 00:53 Why Cursor Trained Composer 2 04:55 Specialization vs Bitter Lesson 06:16 Composer 2 Training Recipe 16:32 Scaling RL Infrastructure Globally 23:32 Floating Point Drift 25:11 MoE Sensitivity Explained 26:25 Router Replay Fix 27:19 Real Time RL Loop 31:49 Long Horizon Agents 34:29 Why RL Everywhere 37:34 LLM as Judge Rewards 39:14 RL in Hard Domains 40:13 Build Your Own Environments 44:34 Closing Thoughts

Today's Training Data episode takes us BTS on the infrastructure challenges required to do large RL runs at scale, featuring Federico Cassano (Composer Lead at Cursor) and Dmytro Dzhulgakov (Co-Founder at Fireworks AI). The Cursor team trained Composer 2 on Fireworks by starting with a strong base model (Kimi 2.5) and performing large-scale mid-training on code tokens and web data to learn common patterns and libraries, followed by a large-scale Reinforcement Learning run to learn how to navigate the Cursor harness, call tools, and write correct code. Today's episode dives into the systems and infrastructure challenges of making that large RL run happening, and there were many (!!), from numerical mismatch to global distribution to synchronizing rollouts across asynchronous pipelines to keeping track of expert activation across runs and more. Extremely nerdy in-the-weeds challenges that Federico and Dima were delighted to nerd out on together :) Beyond RL infra, we also discussed Online vs Simulated rollouts, self-summarization for long-horizon agents, environment design ("the most powerful RL environment is the product itself"), and other technical nuggets. PS: We filmed this episode before the SpaceX news, while the Cursor team was still compute-constrained. While Cursor now has all the flops, the takeaways and hurdles crossed ring true for any serious application-level company that is racing to post-train their own models. I believe that more serious application companies will go the way of Cursor and post-train their own models. 00:00 Introduction 00:53 Why Cursor Trained Composer 2 04:55 Specialization vs Bitter Lesson 06:16 Composer 2 Training Recipe 16:32 Scaling RL Infrastructure Globally 23:32 Floating Point Drift 25:11 MoE Sensitivity Explained 26:25 Router Replay Fix 27:19 Real Time RL Loop 31:49 Long Horizon Agents 34:29 Why RL Everywhere 37:34 LLM as Judge Rewards 39:14 RL in Hard Domains 40:13 Build Your Own Environments 44:34 Closing Thoughts

Sonya Huang 🐥

79,302 views • 2 months ago

The verdict is in! Frontier models can pass the bar, yet they struggle on comprehensive legal research Today we're releasing Legal Research Bench, a benchmark that measures models’ ability to solve realistic legal research tasks across eight areas of U.S. law Instead of awarding partial credit, Legal Research Bench measures whether a model can conduct exhaustive legal analysis. We grade against a strict, all-pass rubric written by practicing lawyers. A model only receives full credit if every required legal element is correct Claude Opus 4.8 leads with 43.8% all-pass accuracy, followed by GPT 5.5 (40.4%) and Claude Sonnet 4.6 (38.5%). While top models score around 80% with partial credit, none exceed 44% when every required legal element must be correct The gap between partial and all-pass accuracy shows how difficult it remains for AI to produce complete, reliable legal research. We hope that Legal Research Bench helps better measure, and ultimately close that gap Lots of exciting work happening in Legal AI from Harvey and Crosby. Excited for the legal research benchmarks ahead!

The verdict is in! Frontier models can pass the bar, yet they struggle on comprehensive legal research Today we're releasing Legal Research Bench, a benchmark that measures models’ ability to solve realistic legal research tasks across eight areas of U.S. law Instead of awarding partial credit, Legal Research Bench measures whether a model can conduct exhaustive legal analysis. We grade against a strict, all-pass rubric written by practicing lawyers. A model only receives full credit if every required legal element is correct Claude Opus 4.8 leads with 43.8% all-pass accuracy, followed by GPT 5.5 (40.4%) and Claude Sonnet 4.6 (38.5%). While top models score around 80% with partial credit, none exceed 44% when every required legal element must be correct The gap between partial and all-pass accuracy shows how difficult it remains for AI to produce complete, reliable legal research. We hope that Legal Research Bench helps better measure, and ultimately close that gap Lots of exciting work happening in Legal AI from Harvey and Crosby. Excited for the legal research benchmarks ahead!

Vals AI

17,186 views • 1 month ago

People are overlooking Google Gemini Realtime models for computer use It gave me sub 100ms latency with computer use. It has a much larger context window and is much cheaper as well Combine that with local OCR and local screen detection model based on Omniparser by Microsoft it works under 100ms action taking when combined with Cua I also put in a harness for Nous Research Hermes with it. You can access it all at your tip of your cursor. You can draw on your screen to give your agents a context And I am making it Open Source! Link in the Comments Sundar Pichai Min-Liang Tan Mojtaba Seyedhosseini

People are overlooking Google Gemini Realtime models for computer use It gave me sub 100ms latency with computer use. It has a much larger context window and is much cheaper as well Combine that with local OCR and local screen detection model based on Omniparser by Microsoft it works under 100ms action taking when combined with Cua I also put in a harness for Nous Research Hermes with it. You can access it all at your tip of your cursor. You can draw on your screen to give your agents a context And I am making it Open Source! Link in the Comments Sundar Pichai Min-Liang Tan Mojtaba Seyedhosseini

Milind S

38,129 views • 1 month ago

Which LLM reasons best when it doesn't have all the information? Enter LLM Poker Arena to find out. It's a Poker Playing benchmark where top reasoning models play Texas Hold'em poker against each other. Claude Opus 4.5, GPT-5.2, Gemini 2.5 Pro, and Grok 4 all sit at the same table and play full tournaments to see who finishes with the chips. Poker is very different when it comes to reasoning. It has to balance probabilistic reasoning, opponent modeling and make decisions under uncertainty. Poker is an interesting evaluation because it tests reasoning under incomplete information, something most coding benchmarks do not capture. In this tournaments the rules are: - Each LLM starts with $1,000 chips - Small and big blinds start at $25 / $50 - Blinds double every 3 minutes - All models run in their reasoning or thinking modes After the first 5 tournaments: - Claude Opus 4.5 with Thinking has 3 wins - GPT-5.2 has 2 wins - Grok 4 and Gemini 2.5 Pro have 0 wins Early results suggest Claude performs quite well at poker as well. Also five is a very small sample size. Planning to run many more tournaments, publish the full benchmark data and add a prediction market on top of it. Thanks for the suggestion clipz. Much more coming as part of Poker Cities !! This was built on Replit ⠕ using their AI integrations, which made it straightforward to connect Claude, GPT, and Gemini. What model do you think wins after 100 tournaments?

Which LLM reasons best when it doesn't have all the information? Enter LLM Poker Arena to find out. It's a Poker Playing benchmark where top reasoning models play Texas Hold'em poker against each other. Claude Opus 4.5, GPT-5.2, Gemini 2.5 Pro, and Grok 4 all sit at the same table and play full tournaments to see who finishes with the chips. Poker is very different when it comes to reasoning. It has to balance probabilistic reasoning, opponent modeling and make decisions under uncertainty. Poker is an interesting evaluation because it tests reasoning under incomplete information, something most coding benchmarks do not capture. In this tournaments the rules are: - Each LLM starts with $1,000 chips - Small and big blinds start at $25 / $50 - Blinds double every 3 minutes - All models run in their reasoning or thinking modes After the first 5 tournaments: - Claude Opus 4.5 with Thinking has 3 wins - GPT-5.2 has 2 wins - Grok 4 and Gemini 2.5 Pro have 0 wins Early results suggest Claude performs quite well at poker as well. Also five is a very small sample size. Planning to run many more tournaments, publish the full benchmark data and add a prediction market on top of it. Thanks for the suggestion clipz. Much more coming as part of Poker Cities !! This was built on Replit ⠕ using their AI integrations, which made it straightforward to connect Claude, GPT, and Gemini. What model do you think wins after 100 tournaments?

Anshul Dhawan

32,192 views • 6 months ago

I got to try Grok 4.5 in early access in Cursor for the past few days and I absolutely enjoyed it. It feels like Opus 4.8 at 2x the speed at a much cheaper price point. I tasked it to brainstorm > plan > implement a big feature for my game (this act 1 boss fight) and it did not disappoint. - It is much smarter than Composer 2.5, during planning mode, it is able to think through my request more robustly, ensuring that edge cases are covered and makes sure to ask the right questions to confirm with me first. - It is much better at brainstorming ideas/suggestions, similar to Opus 4.8, though I think Fable still edges out a little when it comes to brainstorming ideas and suggestions - It is FAST. probably the fastest of all frontier models (Opus 4.8, GPT 5.5 etc), which makes it a joy to build with, because I can stay in the flow - It has much improved visual/animation capabilities than Composer 2.5, it can code up animations (i wanted an explosion animation with particle effects) with much, much better visuals, animation movement and timing. This is a big leap and I was so happy to see this improvement. - The best part for me is that I can just use the same model from planning down to execution without switching to a lower cost model because the price point is cheaper than other frontier models. I'll be testing this model with more challenging tasks in the next few days but I think this is going to be my main driver for vibe coding for a while. Also, its nice to see Grok back in the race. 🙌

I got to try Grok 4.5 in early access in Cursor for the past few days and I absolutely enjoyed it. It feels like Opus 4.8 at 2x the speed at a much cheaper price point. I tasked it to brainstorm > plan > implement a big feature for my game (this act 1 boss fight) and it did not disappoint. - It is much smarter than Composer 2.5, during planning mode, it is able to think through my request more robustly, ensuring that edge cases are covered and makes sure to ask the right questions to confirm with me first. - It is much better at brainstorming ideas/suggestions, similar to Opus 4.8, though I think Fable still edges out a little when it comes to brainstorming ideas and suggestions - It is FAST. probably the fastest of all frontier models (Opus 4.8, GPT 5.5 etc), which makes it a joy to build with, because I can stay in the flow - It has much improved visual/animation capabilities than Composer 2.5, it can code up animations (i wanted an explosion animation with particle effects) with much, much better visuals, animation movement and timing. This is a big leap and I was so happy to see this improvement. - The best part for me is that I can just use the same model from planning down to execution without switching to a lower cost model because the price point is cheaper than other frontier models. I'll be testing this model with more challenging tasks in the next few days but I think this is going to be my main driver for vibe coding for a while. Also, its nice to see Grok back in the race. 🙌

Danny Limanseta

1,410,316 views • 20 days ago

I used Cursor to vibe code a simple fishing game prototype on the Unity Engine. Here's my learnings: - I did not use any Unity MCP for this. The game was built entirely by Cursor models (Sonnet 4.6 for execution and Opus 4.6 for planning) - The model was able to set up the game, getting the basic game mechanics working fairly quickly - I had to use the Unity Game Editor UI to attach components to the in-game objects manually, but it was quite easy to follow the instructions given by the Cursor model - Unity Editor is huge and slow! Compared to Godot, I find the UI really clunky I feel tired looking at it - Unity Assets Marketplace is amazing, there are so many amazing art assets there (like the ones I am using for this game). This is probably the biggest strength of Unity. - I had some issues with restoring checkpoints, probably because of how Unity Game Editor UI being really clunky and I had to manually adjust things in the Editor, which the model doesnt have knowledge of Overall, the results turn out pretty decent, but it was a rather frustrating experience, especially when I had to debug issues or rollback changes. I'll explore more vibe coding on Unity but for now, I think I prefer Godot as a game engine. I just wish there is a Godot Asset Marketplace!

I used Cursor to vibe code a simple fishing game prototype on the Unity Engine. Here's my learnings: - I did not use any Unity MCP for this. The game was built entirely by Cursor models (Sonnet 4.6 for execution and Opus 4.6 for planning) - The model was able to set up the game, getting the basic game mechanics working fairly quickly - I had to use the Unity Game Editor UI to attach components to the in-game objects manually, but it was quite easy to follow the instructions given by the Cursor model - Unity Editor is huge and slow! Compared to Godot, I find the UI really clunky I feel tired looking at it - Unity Assets Marketplace is amazing, there are so many amazing art assets there (like the ones I am using for this game). This is probably the biggest strength of Unity. - I had some issues with restoring checkpoints, probably because of how Unity Game Editor UI being really clunky and I had to manually adjust things in the Editor, which the model doesnt have knowledge of Overall, the results turn out pretty decent, but it was a rather frustrating experience, especially when I had to debug issues or rollback changes. I'll explore more vibe coding on Unity but for now, I think I prefer Godot as a game engine. I just wish there is a Godot Asset Marketplace!

Danny Limanseta

185,200 views • 4 months ago

🚨 CHINA'S AI PLAYBOOK ISN'T TO BEAT OPENAI AND ANTHROPIC. IT'S TO MAKE THEM IRRELEVANT BY MAKING INTELLIGENCE FREE. Chinese labs are giving away frontier-level models for free. Zhipu's GLM-5 already beats Claude Opus 4.5 and Gemini 3.0 Pro on key benchmarks. DeepSeek, Qwen, and Kimi round out 4 of the top 5 open-weight models globally, all released under open licenses anyone can use commercially. The real weapon isn't the models, it's the cost to run them. Chinese provinces like Gansu, Guizhou, and Inner Mongolia are cutting data center power bills by up to 50%. DeepSeek trained its R1 model for around $6 million, compared to roughly $100 million for GPT-4. The remaining blocker is chips. US export controls still cut China off from Nvidia's best hardware. But Zhipu already trained a full frontier model entirely on Huawei's Ascend chips, no Nvidia involved. Huawei's next-gen Atlas 950 cluster, launching late 2026, claims 6.7 times the compute of Nvidia's planned NVL144 systems. If Huawei closes the chip gap while China keeps giving away free, near frontier models running on subsidized power. The entire business model OpenAI and Anthropic are built on, charging hundreds of billions in aggregate revenue for access to intelligence, stops making sense. Why pay premium prices for marginal capability gains when a free alternative running on cheap compute gets you 90% of the way there.

🚨 CHINA'S AI PLAYBOOK ISN'T TO BEAT OPENAI AND ANTHROPIC. IT'S TO MAKE THEM IRRELEVANT BY MAKING INTELLIGENCE FREE. Chinese labs are giving away frontier-level models for free. Zhipu's GLM-5 already beats Claude Opus 4.5 and Gemini 3.0 Pro on key benchmarks. DeepSeek, Qwen, and Kimi round out 4 of the top 5 open-weight models globally, all released under open licenses anyone can use commercially. The real weapon isn't the models, it's the cost to run them. Chinese provinces like Gansu, Guizhou, and Inner Mongolia are cutting data center power bills by up to 50%. DeepSeek trained its R1 model for around $6 million, compared to roughly $100 million for GPT-4. The remaining blocker is chips. US export controls still cut China off from Nvidia's best hardware. But Zhipu already trained a full frontier model entirely on Huawei's Ascend chips, no Nvidia involved. Huawei's next-gen Atlas 950 cluster, launching late 2026, claims 6.7 times the compute of Nvidia's planned NVL144 systems. If Huawei closes the chip gap while China keeps giving away free, near frontier models running on subsidized power. The entire business model OpenAI and Anthropic are built on, charging hundreds of billions in aggregate revenue for access to intelligence, stops making sense. Why pay premium prices for marginal capability gains when a free alternative running on cheap compute gets you 90% of the way there.

The Macro Paper

71,534 views • 28 days ago

BREAKING 🚨: META AI IS PREPARING TO GET NEW AVACADO MODELS, MANUS BROWSER AGENT AND INTEGRATION WITH OPENCLAW! What’s new? - Meta AI website got migrated to the new stack while retaining the same user experience. - New effort selector, email, and calendar connectors are already available to users. - Memory and Projects are in the works. - Avacado and Avacado Thinking models have been spotted in testing. The router still redirects tothe Llama model. - Meta is testing different models from other providers underneath, including Gemini 3 Pro preview, Claude Sonnet 4.5 and GPT-5.2. This part was likely inherited from Manus AI but used only internally. - Scheduled tasks feature is under development. - A new model named Sierra is being tested to power the upcoming browser agent. This likely will be the same agent used in Manus AI currently. - Big Brain mode is in the works, where multiple model responses will be combined into a final answer. - OpenClaw integration has been spotted to let Meta AI connect to your OpenClaw gateway. In short, Meta is on the path to slowly transform into Manus. OpenClaw integration might be big, but something that other Labs may easily adapt as well. We are yet to know if Avacado's performance will surpass recently released models, where Avacado is expected to be released around this spring.

BREAKING 🚨: META AI IS PREPARING TO GET NEW AVACADO MODELS, MANUS BROWSER AGENT AND INTEGRATION WITH OPENCLAW! What’s new? - Meta AI website got migrated to the new stack while retaining the same user experience. - New effort selector, email, and calendar connectors are already available to users. - Memory and Projects are in the works. - Avacado and Avacado Thinking models have been spotted in testing. The router still redirects tothe Llama model. - Meta is testing different models from other providers underneath, including Gemini 3 Pro preview, Claude Sonnet 4.5 and GPT-5.2. This part was likely inherited from Manus AI but used only internally. - Scheduled tasks feature is under development. - A new model named Sierra is being tested to power the upcoming browser agent. This likely will be the same agent used in Manus AI currently. - Big Brain mode is in the works, where multiple model responses will be combined into a final answer. - OpenClaw integration has been spotted to let Meta AI connect to your OpenClaw gateway. In short, Meta is on the path to slowly transform into Manus. OpenClaw integration might be big, but something that other Labs may easily adapt as well. We are yet to know if Avacado's performance will surpass recently released models, where Avacado is expected to be released around this spring.

TestingCatalog News 🗞

188,973 views • 5 months ago

🚨A 25 YEAR OLD BUILT THE FASTEST GROWING SOFTWARE COMPANY IN HISTORY.. WITH ZERO MARKETING SPEND.. AND SPACEX JUST OFFERED $60 BILLION TO BUY IT.. His name is Michael Truell.. He started coding at 11.. Interned at Google at 18.. Dropped out of MIT to start a company that built AI tools for mechanical engineering.. That company failed.. So he pivoted.. And built Cursor.. An AI-powered code editor that writes software for you.. Here's how fast it grew.. $100 million in annual revenue in 12 months.. Fastest in SaaS history.. Broke every record ever set by Slack, Zoom, and Wiz.. $500 million by month 21.. $1 billion by November 2025.. $2 billion by February 2026.. Projected to hit $6 billion by end of year.. Zero marketing spend.. Not a single dollar.. Pure word of mouth from developers who couldn't stop talking about it.. Over 1 billion lines of code accepted per day.. Used by 70% of Fortune 1000 companies.. Every single one of Nvidia's 40,000 engineers uses it.. Coinbase hit 100% adoption among their developers.. And he did this with a team of four MIT co-founders.. One of them was a three-time International Math Olympiad competitor from Pakistan.. Another was a college squash captain with zero startup experience who built the entire product strategy.. They spent zero on sales.. Zero on ads.. Zero on growth hacking.. The product sold itself.. But here's where the story takes a turn nobody expected.. Even at $50 billion valuation.. Even generating billions in revenue.. They hit a wall.. Not a market wall.. A physics wall.. They couldn't get enough GPUs to train their next AI model.. The physical chips didn't exist in sufficient quantities for them to buy.. Money couldn't solve the problem.. Enter Elon Musk.. On April 21.. SpaceX announced a deal to potentially acquire Cursor for $60 billion.. The largest acquisition option in tech history.. The structure is insane.. SpaceX gives Cursor immediate access to Colossus.. xAI's supercomputer equivalent to one million Nvidia H100 GPUs.. For nine months of joint development.. At the end.. SpaceX can buy the company for $60 billion.. If they don't buy it.. They owe Cursor a $10 billion breakup fee.. The largest breakup fee in corporate history.. Think about what that means for Cursor.. Either they get acquired for $60 billion.. Or they walk away with $10 billion in cash and nine months of free training on the most powerful supercomputer on earth.. There is no losing scenario.. And here's why Musk wants it.. SpaceX is preparing for an IPO at $1.75 trillion.. The biggest IPO ever.. But aerospace alone can't justify that number.. By merging xAI into SpaceX.. And now acquiring Cursor.. Musk transforms SpaceX from a rocket company into an AI empire that owns the compute, the models, and the developer tools.. Cursor is the missing piece.. The application layer that puts xAI's models into the daily workflow of every Fortune 500 engineering team.. Oh and one more thing.. In 2022.. FTX's trading firm Alameda Research made a seed investment in Cursor.. During the FTX bankruptcy.. Liquidators sold that stake for $200,000.. That stake is now worth approximately $3 billion.. Sam Bankman-Fried called it the worst liquidation decision in venture capital history.. From a prison cell.. A failed mechanical engineering startup.. Pivoted by four kids from MIT.. Zero marketing.. Zero sales team.. Built the fastest growing software company in history.. And now SpaceX is writing a $60 billion check for it.. This is the most insane founder story in Silicon Valley history.. And most people haven't even heard of Michael Truell.

🚨A 25 YEAR OLD BUILT THE FASTEST GROWING SOFTWARE COMPANY IN HISTORY.. WITH ZERO MARKETING SPEND.. AND SPACEX JUST OFFERED $60 BILLION TO BUY IT.. His name is Michael Truell.. He started coding at 11.. Interned at Google at 18.. Dropped out of MIT to start a company that built AI tools for mechanical engineering.. That company failed.. So he pivoted.. And built Cursor.. An AI-powered code editor that writes software for you.. Here's how fast it grew.. $100 million in annual revenue in 12 months.. Fastest in SaaS history.. Broke every record ever set by Slack, Zoom, and Wiz.. $500 million by month 21.. $1 billion by November 2025.. $2 billion by February 2026.. Projected to hit $6 billion by end of year.. Zero marketing spend.. Not a single dollar.. Pure word of mouth from developers who couldn't stop talking about it.. Over 1 billion lines of code accepted per day.. Used by 70% of Fortune 1000 companies.. Every single one of Nvidia's 40,000 engineers uses it.. Coinbase hit 100% adoption among their developers.. And he did this with a team of four MIT co-founders.. One of them was a three-time International Math Olympiad competitor from Pakistan.. Another was a college squash captain with zero startup experience who built the entire product strategy.. They spent zero on sales.. Zero on ads.. Zero on growth hacking.. The product sold itself.. But here's where the story takes a turn nobody expected.. Even at $50 billion valuation.. Even generating billions in revenue.. They hit a wall.. Not a market wall.. A physics wall.. They couldn't get enough GPUs to train their next AI model.. The physical chips didn't exist in sufficient quantities for them to buy.. Money couldn't solve the problem.. Enter Elon Musk.. On April 21.. SpaceX announced a deal to potentially acquire Cursor for $60 billion.. The largest acquisition option in tech history.. The structure is insane.. SpaceX gives Cursor immediate access to Colossus.. xAI's supercomputer equivalent to one million Nvidia H100 GPUs.. For nine months of joint development.. At the end.. SpaceX can buy the company for $60 billion.. If they don't buy it.. They owe Cursor a $10 billion breakup fee.. The largest breakup fee in corporate history.. Think about what that means for Cursor.. Either they get acquired for $60 billion.. Or they walk away with $10 billion in cash and nine months of free training on the most powerful supercomputer on earth.. There is no losing scenario.. And here's why Musk wants it.. SpaceX is preparing for an IPO at $1.75 trillion.. The biggest IPO ever.. But aerospace alone can't justify that number.. By merging xAI into SpaceX.. And now acquiring Cursor.. Musk transforms SpaceX from a rocket company into an AI empire that owns the compute, the models, and the developer tools.. Cursor is the missing piece.. The application layer that puts xAI's models into the daily workflow of every Fortune 500 engineering team.. Oh and one more thing.. In 2022.. FTX's trading firm Alameda Research made a seed investment in Cursor.. During the FTX bankruptcy.. Liquidators sold that stake for $200,000.. That stake is now worth approximately $3 billion.. Sam Bankman-Fried called it the worst liquidation decision in venture capital history.. From a prison cell.. A failed mechanical engineering startup.. Pivoted by four kids from MIT.. Zero marketing.. Zero sales team.. Built the fastest growing software company in history.. And now SpaceX is writing a $60 billion check for it.. This is the most insane founder story in Silicon Valley history.. And most people haven't even heard of Michael Truell.

Evan Luthra

988,417 views • 3 months ago

The most overlooked part of the SpaceX IPO thesis is the model and most people are completely missing it (Save this) Everyone has been focused on the Anthropic compute deal and the Colossus revenue because those are numbers you can put in a spreadsheet. Six months ago, xAI was competing reasonably well on model performance but was not clearly on the frontier. Then SpaceX exercised its option to acquire Cursor for $60 billion, the largest startup acquisition in history just days after completing the largest IPO in history at $75 billion. Cursor is a team of 700 to 800 people, was on track to exit 2026 at up to $10 billion in revenue, had millions of professional developers using it daily, and had already built a team with the genuine potential to compete at the frontier, the one thing holding them back was compute. SpaceX just gave them the largest GPU cluster in the world to work with. Grok 4.3, a 1.5 trillion parameter model, is currently training with Cursor's proprietary coding data being injected directly into pre-training, not just fine tuning which is a fundamentally more powerful integration than anything the market is currently modeling. The prior version, Grok 4, was already on the Pareto frontier as of 10 to 12 days ago, the most intelligent 500 billion parameter model in the world, sitting alongside Google Gemini, Anthropic, and OpenAI as one of only four systems at the true frontier. Composer 2.5, the previous Cursor model was Pareto dominant in coding tasks just before the acquisition closed, meaning SpaceX inherited a model that was already best-in-class in the highest-value AI use case in the market. The AWS parallel is the one everyone keeps missing. Bezos built data center capacity for Black Friday, sat on idle infrastructure the rest of the year, and monetized it into what was at the time the most profitable technology business in history and investors hated it in 2009 and 2010 because he was burning free cash flow on capacity that had no obvious revenue yet. SpaceX is in exactly that position, it built Colossus for xAI's own training needs, is monetizing excess capacity to Anthropic at $1.25 billion per month across 220,000 Nvidia GPUs, and has reportedly secured up to 20% of Nvidia's early Vera Rubin allocation, giving it the most powerful and scarcest GPU infrastructure in the world during the critical window when those chips are hardest to get. The $60 billion Cursor acquisition closed at a moment when SpaceX had essentially unlimited compute, a team already at the frontier, and a product with deep enterprise distribution, three things no other model lab had simultaneously when it was at this stage. The market is pricing the compute business conservatively and ignoring the model call option entirely, and coding is the fastest path to AGI, once you are on the Pareto frontier with that compute, revenue scales fast. Anthropic went from negligible revenue to $30 billion annualized in under 18 months and that is the existence proof. Bullish on SpaceXAI and Elon Musk

The most overlooked part of the SpaceX IPO thesis is the model and most people are completely missing it (Save this) Everyone has been focused on the Anthropic compute deal and the Colossus revenue because those are numbers you can put in a spreadsheet. Six months ago, xAI was competing reasonably well on model performance but was not clearly on the frontier. Then SpaceX exercised its option to acquire Cursor for $60 billion, the largest startup acquisition in history just days after completing the largest IPO in history at $75 billion. Cursor is a team of 700 to 800 people, was on track to exit 2026 at up to $10 billion in revenue, had millions of professional developers using it daily, and had already built a team with the genuine potential to compete at the frontier, the one thing holding them back was compute. SpaceX just gave them the largest GPU cluster in the world to work with. Grok 4.3, a 1.5 trillion parameter model, is currently training with Cursor's proprietary coding data being injected directly into pre-training, not just fine tuning which is a fundamentally more powerful integration than anything the market is currently modeling. The prior version, Grok 4, was already on the Pareto frontier as of 10 to 12 days ago, the most intelligent 500 billion parameter model in the world, sitting alongside Google Gemini, Anthropic, and OpenAI as one of only four systems at the true frontier. Composer 2.5, the previous Cursor model was Pareto dominant in coding tasks just before the acquisition closed, meaning SpaceX inherited a model that was already best-in-class in the highest-value AI use case in the market. The AWS parallel is the one everyone keeps missing. Bezos built data center capacity for Black Friday, sat on idle infrastructure the rest of the year, and monetized it into what was at the time the most profitable technology business in history and investors hated it in 2009 and 2010 because he was burning free cash flow on capacity that had no obvious revenue yet. SpaceX is in exactly that position, it built Colossus for xAI's own training needs, is monetizing excess capacity to Anthropic at $1.25 billion per month across 220,000 Nvidia GPUs, and has reportedly secured up to 20% of Nvidia's early Vera Rubin allocation, giving it the most powerful and scarcest GPU infrastructure in the world during the critical window when those chips are hardest to get. The $60 billion Cursor acquisition closed at a moment when SpaceX had essentially unlimited compute, a team already at the frontier, and a product with deep enterprise distribution, three things no other model lab had simultaneously when it was at this stage. The market is pricing the compute business conservatively and ignoring the model call option entirely, and coding is the fastest path to AGI, once you are on the Pareto frontier with that compute, revenue scales fast. Anthropic went from negligible revenue to $30 billion annualized in under 18 months and that is the existence proof. Bullish on SpaceXAI and Elon Musk

Milk Road AI

69,446 views • 1 month ago

We are excited to share that Logan Kilpatrick joined us on The Bench to discuss Google's new Gemini 3.5 Flash: why it's deliberately more persistent and capable than previous Flash models, how it hit #1 on our FinanceAgent Benchmark taking 82 steps where competitors stopped at 13, and what justifies the price increase. We also get into why AI benchmarks need a paradigm shift, the trade-off of building everything vs staying focused, the Pope, and why Omni might kill the Subway Surfers content era. 0:11:00 – Flash is being rebased for the agent era, not just a cheaper model anymore 0:14:03 – Persistence by design: 82 tool calls vs competitors' 13 0:17:52 – Why pricing went up and how Google thinks about value per token 0:22:55 – Coding performance: from 20th to 10th place in one generation 0:28:28 – Why benchmarks have historically been misleading and what the new era of evaluation looks like 0:29:28 Logan on why Google has the best researchers in the world 0:36:16 – The cost of being Google 0:39:07 – The Pope’s encyclical on AI and whether most people see frontier intelligence as a good thing 0:51:12 – Why Omni is the thing that recently clicked

We are excited to share that Logan Kilpatrick joined us on The Bench to discuss Google's new Gemini 3.5 Flash: why it's deliberately more persistent and capable than previous Flash models, how it hit #1 on our FinanceAgent Benchmark taking 82 steps where competitors stopped at 13, and what justifies the price increase. We also get into why AI benchmarks need a paradigm shift, the trade-off of building everything vs staying focused, the Pope, and why Omni might kill the Subway Surfers content era. 0:11:00 – Flash is being rebased for the agent era, not just a cheaper model anymore 0:14:03 – Persistence by design: 82 tool calls vs competitors' 13 0:17:52 – Why pricing went up and how Google thinks about value per token 0:22:55 – Coding performance: from 20th to 10th place in one generation 0:28:28 – Why benchmarks have historically been misleading and what the new era of evaluation looks like 0:29:28 Logan on why Google has the best researchers in the world 0:36:16 – The cost of being Google 0:39:07 – The Pope’s encyclical on AI and whether most people see frontier intelligence as a good thing 0:51:12 – Why Omni is the thing that recently clicked

Vals AI

16,486 views • 2 months ago

China just released an open source AI model that matches the best closed models from OpenAI and Anthropic. Gavin Baker explained exactly how they did it and the answer should concern every American AI lab. The model is called GLM 5.2. It was built by Z. AI. You get 744 billion parameters, 1 million token context window and its MIT license, meaning anyone can download it, fork it, build a company on it, with no restrictions and no Dario. It scored 51 points on the artificial analysis intelligence index. The highest score any open weight model has ever achieved. It beat GPT 5.5 on the frontier software engineering benchmark. It trails Claude Opus 4.8 by less than one percentage point. And it costs 85% less to run than GPT 5.5 for comparable performance. Gavin Baker said on the All-In podcast that this model has challenged some of his beliefs. Then he explained how China built it. The method is called distillation. Just think of tens of thousands of phones and computers running simultaneously, all hitting the frontier model APIs through masked accounts, asking specific questions, and harvesting what happens inside the model when it answers. Every reasoning step, every token. The entire thinking process gets recorded and fed back into the Chinese model during training. It is a cheat sheet. It is the answer key to the exam. And here is the part that should worry everyone. Sacks said it plainly. China was already nine months behind American models. But now that GLM 5.2 is good enough to run its own reinforcement learning, it can improve itself without needing to distill from American models anymore. The cheat sheet let them get close enough to start writing their own answers. Sacks said we are six months behind on the model and 24 months behind on silicon and they are only a few months behind in total. The Z. AI founder told Elon Musk directly that open weight fable-level capability will be here before Q1 2027. Every restriction Anthropic lobbied for, every self-imposed safety guardrail, every month of delay in releasing American frontier models accelerated this. The Chinese labs were not under those restrictions. They were not going to wait. The composable model future Gavin described, where every enterprise runs a frontier model alongside their own fine-tuned open weight model, is coming regardless of what American labs do next. The question is just whether the open weight half of that stack is American or Chinese. Right now it is Chinese. WATCH THE FULL PODCAST ON The All-In Podcast

China just released an open source AI model that matches the best closed models from OpenAI and Anthropic. Gavin Baker explained exactly how they did it and the answer should concern every American AI lab. The model is called GLM 5.2. It was built by Z. AI. You get 744 billion parameters, 1 million token context window and its MIT license, meaning anyone can download it, fork it, build a company on it, with no restrictions and no Dario. It scored 51 points on the artificial analysis intelligence index. The highest score any open weight model has ever achieved. It beat GPT 5.5 on the frontier software engineering benchmark. It trails Claude Opus 4.8 by less than one percentage point. And it costs 85% less to run than GPT 5.5 for comparable performance. Gavin Baker said on the All-In podcast that this model has challenged some of his beliefs. Then he explained how China built it. The method is called distillation. Just think of tens of thousands of phones and computers running simultaneously, all hitting the frontier model APIs through masked accounts, asking specific questions, and harvesting what happens inside the model when it answers. Every reasoning step, every token. The entire thinking process gets recorded and fed back into the Chinese model during training. It is a cheat sheet. It is the answer key to the exam. And here is the part that should worry everyone. Sacks said it plainly. China was already nine months behind American models. But now that GLM 5.2 is good enough to run its own reinforcement learning, it can improve itself without needing to distill from American models anymore. The cheat sheet let them get close enough to start writing their own answers. Sacks said we are six months behind on the model and 24 months behind on silicon and they are only a few months behind in total. The Z. AI founder told Elon Musk directly that open weight fable-level capability will be here before Q1 2027. Every restriction Anthropic lobbied for, every self-imposed safety guardrail, every month of delay in releasing American frontier models accelerated this. The Chinese labs were not under those restrictions. They were not going to wait. The composable model future Gavin described, where every enterprise runs a frontier model alongside their own fine-tuned open weight model, is coming regardless of what American labs do next. The question is just whether the open weight half of that stack is American or Chinese. Right now it is Chinese. WATCH THE FULL PODCAST ON The All-In Podcast

Ihtesham Ali

86,163 views • 1 month ago

Chamath is making one of the most important business arguments of 2026. Half of large US companies right now cannot generate returns that exceed their cost of capital, which has normalized back to its long run average of 8 to 11%. Another one in seven companies globally is stuck generating persistent returns between 1 and 5% and most businesses don't have room for error and in this environment walks every frontier AI lab saying the same thing, give us your data, your workflows, your processes and our model will make everything better. And companies by the millions said yes. What they didn't fully account for is what happens on the other side of that door. Every time an employee runs a query through a frontier model API, the prompt goes through external servers, workflows, customer data, pricing logic, internal processes, all of it transmitted through a third party. As Alex Karp said companies are spending on tokens while handing over the exact proprietary advantages that make their business worth owning. Microsoft blocked internal use of Anthropic's Claude Fable 5 but over its 30-day data retention policy and the largest software company in the world decided a frontier model's data handling was too risky for its own employees. A US government action revoked access to another frontier model for foreign nationals overnight. Now here's where the cost math becomes impossible to ignore. Deutsche Bank calculated a roughly 65x cost gap between frontier models like Claude Fable 5 at ~$3.25 per task and open-source alternatives at ~$0.05. For 90% of everyday enterprise tasks, performance is comparable. Open-weight models now match closed frontier systems on core agent tasks at roughly one-tenth the cost, a high-volume deployment that costs $250/day on Claude runs at $12/day on an open-source equivalent. Chamath Palihapitiya tested this directly by running a standard enterprise code migration task through an orchestration layer wrapping an open-source model came in 16.4x cheaper than using a frontier model directly.

Chamath is making one of the most important business arguments of 2026. Half of large US companies right now cannot generate returns that exceed their cost of capital, which has normalized back to its long run average of 8 to 11%. Another one in seven companies globally is stuck generating persistent returns between 1 and 5% and most businesses don't have room for error and in this environment walks every frontier AI lab saying the same thing, give us your data, your workflows, your processes and our model will make everything better. And companies by the millions said yes. What they didn't fully account for is what happens on the other side of that door. Every time an employee runs a query through a frontier model API, the prompt goes through external servers, workflows, customer data, pricing logic, internal processes, all of it transmitted through a third party. As Alex Karp said companies are spending on tokens while handing over the exact proprietary advantages that make their business worth owning. Microsoft blocked internal use of Anthropic's Claude Fable 5 but over its 30-day data retention policy and the largest software company in the world decided a frontier model's data handling was too risky for its own employees. A US government action revoked access to another frontier model for foreign nationals overnight. Now here's where the cost math becomes impossible to ignore. Deutsche Bank calculated a roughly 65x cost gap between frontier models like Claude Fable 5 at ~$3.25 per task and open-source alternatives at ~$0.05. For 90% of everyday enterprise tasks, performance is comparable. Open-weight models now match closed frontier systems on core agent tasks at roughly one-tenth the cost, a high-volume deployment that costs $250/day on Claude runs at $12/day on an open-source equivalent. Chamath Palihapitiya tested this directly by running a standard enterprise code migration task through an orchestration layer wrapping an open-source model came in 16.4x cheaper than using a frontier model directly.

Milk Road AI

280,929 views • 24 days ago

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.

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

116,971 views • 7 days ago

Aman Sanger on scaling Cursor to $100M in revenue in 12 months The four co-founders started working on Cursor in January 2023. “It was about 2-4 months of experimenting, trying different things and finding something that fit,” Aman recalls. “Then we shipped it and it had an okay launch.” The initial version of Cursor got some initial buzz because they shipped it with GPT-4 at a time where very few products were using Open AI’s latest model. But then usage tanked. The team began to doubt their approach after the failed launch. Aman explains: “The entirety of that summer was just incredibly slow growth, and that was somewhat demoralizing. The big question in our minds was, ‘Are we being too ambitious?’ We were trying to build this general purpose thing for all engineers, but with this really small team, maybe we should focus more narrowly on some particular use case like tests or bug detection.” But the fact that they were users of the product gave the team the confidence to keep going down their initial path: “The really magical thing about this product was that we were users of it. So we could iterate incredibly quickly, and we tried all these different things that summer. Then we found this core set of features that worked incredibly well.” The two key features were Command K for instructed edit ability and code-based indexing that let you ask questions about your whole code base. “After we integrated those two features and launched them, growth kind of just took off.” Aman continues: “A lot of the work of Cursor has been just experimenting with what is possible. For everything you see in the product, there’s like 10 failed experiments that didn’t work… All of our work for the first 6 months to a year was trying to find new ways to harness these models and make these models better for programming.” Another factor that contributed to Cursor’s success was their willingness to ship half-finished features: “We released these half-finished things, which a lot of our competitors refused to do. The first version of Copilot++ and Cursor Tab sucked. But once you release it to the world and see how people react to it, you can improve on it a ton… We biased toward releasing as soon as something shows signs of usefulness to the team.” Video source: Peak XV Partners (2025)

Aman Sanger on scaling Cursor to $100M in revenue in 12 months The four co-founders started working on Cursor in January 2023. “It was about 2-4 months of experimenting, trying different things and finding something that fit,” Aman recalls. “Then we shipped it and it had an okay launch.” The initial version of Cursor got some initial buzz because they shipped it with GPT-4 at a time where very few products were using Open AI’s latest model. But then usage tanked. The team began to doubt their approach after the failed launch. Aman explains: “The entirety of that summer was just incredibly slow growth, and that was somewhat demoralizing. The big question in our minds was, ‘Are we being too ambitious?’ We were trying to build this general purpose thing for all engineers, but with this really small team, maybe we should focus more narrowly on some particular use case like tests or bug detection.” But the fact that they were users of the product gave the team the confidence to keep going down their initial path: “The really magical thing about this product was that we were users of it. So we could iterate incredibly quickly, and we tried all these different things that summer. Then we found this core set of features that worked incredibly well.” The two key features were Command K for instructed edit ability and code-based indexing that let you ask questions about your whole code base. “After we integrated those two features and launched them, growth kind of just took off.” Aman continues: “A lot of the work of Cursor has been just experimenting with what is possible. For everything you see in the product, there’s like 10 failed experiments that didn’t work… All of our work for the first 6 months to a year was trying to find new ways to harness these models and make these models better for programming.” Another factor that contributed to Cursor’s success was their willingness to ship half-finished features: “We released these half-finished things, which a lot of our competitors refused to do. The first version of Copilot++ and Cursor Tab sucked. But once you release it to the world and see how people react to it, you can improve on it a ton… We biased toward releasing as soon as something shows signs of usefulness to the team.” Video source: Peak XV Partners (2025)

Startup Archive

56,166 views • 1 year ago

AgentRouter is handing out $125 in FREE API credits for OPUS 4.8 > No card. > GitHub sign-in. > One API key for Opus 4.8 / Opus 4.7 / Sonnet 4.6 / GLM 5.1 The useful part is not “another model gateway.” It is that you can point coding tools and agents at one OpenAI-compatible base URL instead of juggling separate keys for Claude and GLM, and others. What you get: • $125 signup balance • one API key • OpenAI-compatible endpoints • chat, responses, messages, embeddings, images, audio, rerank endpoints listed • claimed support for Cursor, Claude Code, Hermes, and other agent tools The flow is simple: 1. go to AgentRouter (link in comment) 2. sign in or create an account 3. generate an API key in the console 4. copy the base URL 5. paste it into Cursor / Claude Code / your agent runtime 6. run a small test before moving anything serious Caveat: AgentRouter’s own site currently has a notice saying Claude-series service was recently hit by stability issues and large-scale Claude access was suspended. So don’t treat this as “guaranteed free Claude forever.” Treat it as $125 of free routing credits to test which models are live, how the latency feels, and whether the gateway is stable enough for your workflow. Still a very solid save if you build with agents. Most people burn paid API credits just testing configs. This gives you room to test first, then decide if it belongs in your stack.

AgentRouter is handing out $125 in FREE API credits for OPUS 4.8 > No card. > GitHub sign-in. > One API key for Opus 4.8 / Opus 4.7 / Sonnet 4.6 / GLM 5.1 The useful part is not “another model gateway.” It is that you can point coding tools and agents at one OpenAI-compatible base URL instead of juggling separate keys for Claude and GLM, and others. What you get: • $125 signup balance • one API key • OpenAI-compatible endpoints • chat, responses, messages, embeddings, images, audio, rerank endpoints listed • claimed support for Cursor, Claude Code, Hermes, and other agent tools The flow is simple: 1. go to AgentRouter (link in comment) 2. sign in or create an account 3. generate an API key in the console 4. copy the base URL 5. paste it into Cursor / Claude Code / your agent runtime 6. run a small test before moving anything serious Caveat: AgentRouter’s own site currently has a notice saying Claude-series service was recently hit by stability issues and large-scale Claude access was suspended. So don’t treat this as “guaranteed free Claude forever.” Treat it as $125 of free routing credits to test which models are live, how the latency feels, and whether the gateway is stable enough for your workflow. Still a very solid save if you build with agents. Most people burn paid API credits just testing configs. This gives you room to test first, then decide if it belongs in your stack.

kocer

24,905 views • 1 month ago

Cerebras inference is very fast. So fast that it changes how we think about configuring our LLMs for voice agent use cases. Kimi K2.6 is a 1T parameter reasoning model that Cerebras serves at 650 - 1,000 tokens per second (end-to-end throughput), with time to first token metrics as low as 150ms (latency). These numbers are two to three times faster than other similarly capable models. The biggest lever we get from this kind of speed is that we can use the model in reasoning mode, and still have excellent "time to first non-thinking token." This solves a big pain point we have in 2026 for voice agent use cases. Almost all recent innovation in post-training has focused on making models good at reasoning ("test time compute"). This is great, but it makes the user-facing model latency much, much slower. Which is a problem for conversational voice agents. We can run Kimi K2.6 with reasoning turned on, and get responses faster than other models produce with reasoning disabled. On my 30-turn voice agent benchmark, Kimi K2.6 with reasoning enabled ties GPT 5.1 and Haiku 4.5 with reasoning disabled, and is still about 200ms seconds faster! On my primary task agent benchmark, Kimi K2.6 is now the #2 model. It ranks just behind Gemini 3.5 Flash in "high" reasoning mode, and tied with GLM 5, Sonnet 4.6, and GPT 5.4 with reasoning set to "low." But Kimi K2.6 completes each turn in the agent loop in under 500ms. The other four models are all at least 3x slower. (Models only qualify for this benchmark if they can complete task turns at a P50 <4s.) A couple of other things that this speed buys us, for production voice agents: - Tool calls happen fast enough that we don't have to work around tool call latency in our pipeline design. - We can prompt the model to output structured data at the beginning of a response, followed by plain text for voice generation. This opens up possibilities like asking the model to do complex classification/generation tasks that influence the rest of the pipeline. For example, the model could create a detailed style prompt for a steerable TTS model, for each individual conversation turn. And, of course, you can use Kimi K2.6 with reasoning turned off. Cerebras calls this "instant" mode. Here's a video of a Cerebras Kimi K2.6 voice agent with voice-to-voice response time, measured at the client, under 500ms. This is the true response latency as perceived by the user, including all network and audio codec overhead, transcription and turn detection, Kimi K2.6 token generation, and voice generation. 500ms is, effectively, instant. So the Cerebras naming for this mode is a propos. :-)

Cerebras inference is very fast. So fast that it changes how we think about configuring our LLMs for voice agent use cases. Kimi K2.6 is a 1T parameter reasoning model that Cerebras serves at 650 - 1,000 tokens per second (end-to-end throughput), with time to first token metrics as low as 150ms (latency). These numbers are two to three times faster than other similarly capable models. The biggest lever we get from this kind of speed is that we can use the model in reasoning mode, and still have excellent "time to first non-thinking token." This solves a big pain point we have in 2026 for voice agent use cases. Almost all recent innovation in post-training has focused on making models good at reasoning ("test time compute"). This is great, but it makes the user-facing model latency much, much slower. Which is a problem for conversational voice agents. We can run Kimi K2.6 with reasoning turned on, and get responses faster than other models produce with reasoning disabled. On my 30-turn voice agent benchmark, Kimi K2.6 with reasoning enabled ties GPT 5.1 and Haiku 4.5 with reasoning disabled, and is still about 200ms seconds faster! On my primary task agent benchmark, Kimi K2.6 is now the #2 model. It ranks just behind Gemini 3.5 Flash in "high" reasoning mode, and tied with GLM 5, Sonnet 4.6, and GPT 5.4 with reasoning set to "low." But Kimi K2.6 completes each turn in the agent loop in under 500ms. The other four models are all at least 3x slower. (Models only qualify for this benchmark if they can complete task turns at a P50 <4s.) A couple of other things that this speed buys us, for production voice agents: - Tool calls happen fast enough that we don't have to work around tool call latency in our pipeline design. - We can prompt the model to output structured data at the beginning of a response, followed by plain text for voice generation. This opens up possibilities like asking the model to do complex classification/generation tasks that influence the rest of the pipeline. For example, the model could create a detailed style prompt for a steerable TTS model, for each individual conversation turn. And, of course, you can use Kimi K2.6 with reasoning turned off. Cerebras calls this "instant" mode. Here's a video of a Cerebras Kimi K2.6 voice agent with voice-to-voice response time, measured at the client, under 500ms. This is the true response latency as perceived by the user, including all network and audio codec overhead, transcription and turn detection, Kimi K2.6 token generation, and voice generation. 500ms is, effectively, instant. So the Cerebras naming for this mode is a propos. :-)

kwindla

40,319 views • 2 months ago