Todd Cohen retweeted
Video editing is following the same path AI coding took in 2025–26.
The 10x–100x programmers don’t write code anymore. They build software factories.
The people who deeply understand video, brand and storytelling will soon be building content factories.
We got into @ycombinator to fix how AI makes music.
Today we're launching Taan v1-preview, the most controllable foundation music model.
Music models today make catchy tracks, but you can't steer them. Ask for "violin at exactly the 11th second" and they usually skip it. So you regenerate over and over and hope one take fits.
Taan is trained to fix this.
Hum a melody, name exact notes, or hand it an image, a MIDI file, a video or a script. Taan composes music that fits your context and follows your instructions.
v1-preview has genre support limited to orchestral and cinematic music, with more genres on the way.
live today 👇
Todd Cohen retweeted
The CEO of Vercel called us out. We claimed Agent37.com was 25x cheaper, and he said we didn't consider active CPU time.
So we fixed the math..
We gave 100 coding agents a sandbox each and had them vibe code an app. Here's the cost per hour at just 5% active CPU:
Vercel: $9.75 (14x)
E2B: $16.56 (24x)
Daytona: $16.56 (24x)
Agent37: $0.68
We're starting a race to the bottom... see you there.
Find a cheaper sandbox and we'll refund the difference.
Replying to @paarangatrai
WAIT
a total miscalculation of active cpu time¹ to produce fake cost comparisons and fake impressions!?!?
¹ vercel.com/docs/sandbox/pric…
Todd Cohen retweeted
We just killed OpenRouter.
AI routers made $100+ million last year charging 5% markups on every dollar.
Today that ends.
Introducing Straitly the world’s first AI gateway that charges 0% markup on your tokens. No Stripe or Credit Card processing fees.
With one API key, you can access 142+ models from than 20+ providers.
Straitly auto handles handles provider failures via intelligent re-routing achieving a 99.98% success rate while still optimizing for cache hit rates.
To celebrate our launch we are giving up to 30% off your first $10k (yes, $10,000) of token spend, plus $100 in free credits upfront.
If you are an ambitious builder who's spending lots on tokens we want to hear from you!
Comment below what you're building + highest usage model, then head over to straitly.ai and fill out a form with your @x username.
1. The State Machine Hunter
"Find all of the implicit state machines in my app. Make the best decision that will elevate all three: the user experience (UX), developer experience (DX), and agent experience (AX) always, without breaking anything."
Todd Cohen retweeted
here's how to make talking head videos with seed dance 2.5 for about $4
1. write you script in a hook + desired outcome format, my product is bridge to this life
2. have an AI voice do the ad read of the script and make an MP3 file
3. then you send this mp3 as audio reference + the script + a reference image to seed dance 2.5
4. generate it as 480p, then upscale using ffmpeg and edit using word by word transcription to confirm the ad read was correct, any errors swap the mp3 source with the mp4 final error, like you editing out the mistake
5. generate 20 ad variations for $100
6. Upload all of these to facebook ads API
7. Manage the account with the same API key, do your data analytics across your whole stack by putting your FB ads, google analytics, CRM, and payment processor or whatever else
and I open sourced this whole thing
you can modify it with your claude code to ad features, or run the whole software using your coding agent and never touch the UI
get started for free - graphed.com/
here is the skill - github.com/graphed-inc/marke…
Todd Cohen retweeted
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery!
Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window.
Highlights:
- Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation.
- Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables.
- Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery.
- $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%.
The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder.
Start building today! 🚀
- Blog: qwen.ai/blog?id=qwen4-max
- API: qwencloud.com/models/qwen4-m…
- Token Plan: qwencloud.com/pricing/token-…
More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
Fable 5.2 + Opus 5.5 is 𝘀𝗰𝗮𝗿𝘆 𝗴𝗼𝗼𝗱 𝗮𝘁 𝗨𝗜. Before it touches code, have it run your app, 𝘀𝗰𝗿𝗲𝗲𝗻𝘀𝗵𝗼𝘁 𝗲𝘃𝗲𝗿𝘆 𝗽𝗮𝗴𝗲 𝗮𝗻𝗱 𝗽𝗶𝗰𝗸 𝗶𝘁 𝗮𝗽𝗮𝗿𝘁. This works really well.
Launch claude --chrome, run /advisor fable (Opus builds, Fable reviews), then paste:
"Run my app, open every page in Chrome yourself and screenshot it. List the 5 biggest hierarchy and density problems first. Then define a spacing scale, type scale and max 5 color tokens. Rebuild using only those. No gradients. Screenshot at 1440 and 390, fix what looks off"
Todd Cohen retweeted
由 OpenAI 董事会主席兼前 Salesforce 联席 CEO Bret Taylor,联手前 Google Labs 掌门人 Clay Bavor 创立的百亿美金智能体独角兽 Sierra,刚刚正式发布 Personal Agent Protocol v1.0,简称 Poppy。
只用了一个月,Meta、Shopify、Stripe、Salesforce,甚至大家熟悉的 Manus、Browserbase 等 100 多家巨头和明星创企全部入局。
下面总结 5 条关键内容和案例:
1、以前个人智能体帮人办事,只能模拟人类去截图,看网页,点鼠标,又慢又容易崩。Poppy v1.0 相当于给所有网站立了一个智能体专用前台,直接把 MCP、OpenAPI 和 UCP 串成统一标准,让智能体一进门就能读取结构化能力清单,同时底层采用 OAuth 鉴权,用户随时可控权限,遇到付钱、转账等高危操作强制触发人类审批。
2、协议最怕没人用,但 Poppy 一上来就打通了供需两端。在流量侧,Meta 的个人智能体 Muse 已默认全量开启,访问任何网站第一件事就是自动寻找 Poppy 文件。在供给侧,覆盖 175 个国家的 Shopify 已为平台上所有商家自动生成并发布 Poppy 文件,全球数百万独立站无需写一行代码,直接具备接待个人智能体的能力。
3、Shopify 与 Sierra 联合编写了商务扩展,把智能体购物的全链路彻底打通,Poppy 负责进门握手(声明智能体身份、选游客模式还是登录账号、敏感操作前向用户讨要授权),UCP(通用商务协议)负责交易执行(搜商品、加购、结账),遇到复杂导购问题,还能通过 UCP Ask 让你的个人智能体,直接和店铺的客服智能体(A2A)对话。
4、不止电商,高风控行业已经实装上线,在硅谷金融平台 Mercury,个人智能体能直接查余额、给支出分类,甚至发起转账(底层强制走账户预设的审批流才会打款)。在土耳其航空,智能体能查票、用预订号调行程,还能帮会员盯商务舱里程升舱名额并推给用户一键确认,消费金融巨头 Synchrony 则支持智能体在跨 50 万家门店导购时,直接匹配免息分期和返现优惠。
5、最新加入的 100+ 合作伙伴名单中,除了传统金融和航司,还有 Salesforce(含 Agentforce)、Cisco、Zoom、Twilio,以及 AI 圈大家都熟悉的 Manus、Browserbase、WorkOS。
当日常任务全面交由个人智能体代劳,未来不能被智能体通过标准协议丝滑调用的网站,就像移动互联网时代没做手机端适配一样,会在无声无息中丢掉所有自然流量。
Todd Cohen retweeted
You'd rather write code than do the marketing outreach you're supposes go be doing.
Use this playbook. You get to use AI a lot so you'll like it, maybe.
C’mon, go get those customers.
For the first time ever, I'm sharing everything we do across our B2B clients to generate millions of LinkedIn impressions and turn them into qualified meetings and revenue
Every LinkedIn GTM lesson I've learned after generating 50M+ views, 300,000+ followers and $8M+ in tracked client revenue is here:
1. We start by researching the entire market before writing a single post
We track 20–30 competitors, industry creators and adjacent companies targeting the same ICP
Using Apify and GrokBot, we monitor their best-performing posts, identify content outliers and track the topics generating the most attention
Then we analyze the actual ideas behind those posts rather than just copying their hooks
2. We build a content strategy around the specific people our clients want to close
For a B2B SaaS, that might mean founders, CROs, heads of growth and RevOps leaders
We map their problems, buying triggers, objections and the types of content they're most likely to find useful
Every post needs a reason to exist beyond generating impressions
3. We publish 30+ LinkedIn posts per month for each client
But we don't just write generic educational content
We split everything across four categories:
→ Distribution posts to reach new audiences
→ Authority posts demonstrating expertise and real results
→ Lead magnets to capture demand
→ Conversion posts that help prospects understand the offer
The mix changes depending on the client, their market and what the performance data tells us
4. We turn our clients' internal knowledge into content their competitors can't replicate
Sales calls, customer interviews, case studies, internal experiments, product updates, founder opinions and operational workflows
We use Fathom transcripts, internal documents and regular founder interviews to extract the most interesting information
Then we turn those insights into posts with actual examples, numbers and experiences
This is one of the biggest reasons our content works
5. We build 8 high-value lead magnets every month
We create complete AI brains, playbooks, templates, databases, calculators, frameworks and workflows that solve specific problems for the ICP
For example:
→ AI SaaS: Complete AI agent implementation workflows
→ Sales software: Prospecting systems and outbound templates
→ Agencies: Acquisition playbooks and campaign audit frameworks
→ Enterprise services: ROI calculators and implementation roadmaps
We also create detailed visuals showing exactly what's inside each resource
The goal is to make the value obvious before someone even comments
6. We automate the entire lead magnet funnel
Using LeadShark, we connect the LinkedIn post to automated DM delivery and an opt-in page
Someone comments for a resource, receives it automatically and can enter the client's CRM after submitting their details
We can then connect the opt-in to email sequences and relevant follow-up content
The important part is that nobody on the client's team needs to manually deliver hundreds of resources
7. We qualify the people engaging with our content
A post generating 2,000 comments doesn't mean 2,000 potential customers
We use Traxy to monitor engagement across our clients' posts and their competitors', then qualify those leads against the ICP
Company size, industry, job title, seniority and buying intent
Then we separate high-value decision-makers from everyone else
8. We send qualified lead notifications directly into Slack or Telegram
When a relevant founder, executive or decision-maker requests a resource, the sales team can receive a notification with their details and context
This means they don't need to scroll through hundreds of comments looking for potential opportunities
They can focus on the people who actually match the companies they want to close
9. We build warm outbound systems around LinkedIn engagement
Instead of relying entirely on inbound meetings, we use Traxy to spot relevant people interacting with the client's content, competitor posts, or resource giveaways
Then we research their companies and run personalized LinkedIn and email follow-ups based on the exact post or signal they engaged with
Someone who requested a revenue operations framework should receive a different message from someone who engaged with a founder story
The outreach needs to be relevant to what the person actually showed interest in
10. We optimize every client's LinkedIn profile for conversion
Because generating 100,000 impressions means very little if someone visits the founder's profile and has no idea what the company does
We optimize:
→ Headline and positioning
→ Banner and CTA
→ Featured section
→ Case studies and social proof
→ About section
→ Booking links and conversion paths
The profile should work like a landing page for the founder's expertise and the company's offer
11. We build email nurture sequences for people who aren't ready to buy
Not everyone requesting a lead magnet wants to book a sales call immediately
Some people need to see more customer results, understand the problem better or wait until the timing makes sense
We use email sequences to distribute relevant case studies, educational content, implementation examples and opportunities to learn more
The objective is to stay relevant without forcing every lead into an immediate sales conversation
12. We monitor performance and change the strategy based on actual results
Every week, we track impressions, engagement, profile views, relevant followers, resource requests and qualified conversations
Then we connect those numbers to booked meetings, attended calls, opportunities and revenue wherever attribution is available
A post generating 500,000 impressions can be less valuable than a case study reaching 10,000 people if the second attracts the exact buyers we're trying to close
We care about distribution, but we care even more about what that distribution produces
13. We scale the system across multiple executive accounts
Once the founder's content is working, we can expand to the CRO, head of growth, CTO or other relevant people inside the company
Each person gets a distinct audience and content strategy
For example:
→ CEO: Founders and executives
→ CRO: Revenue leaders
→ Head of Growth: Marketers and acquisition teams
→ CTO: Technical buyers
→ Customer Success: Practitioners and implementation teams
All the accounts can share the same research, content production and reporting infrastructure while publishing from their own expertise
We've seen clients expand from a single founder account into company-wide distribution networks
14. We turn the best-performing content into assets for the entire business
A successful LinkedIn post can become a YouTube video, newsletter, sales enablement resource, outbound follow-up or lead magnet
A customer case study can become five different pieces of content for five different buyer personas
And a useful internal workflow can become a public playbook that attracts potential customers
We want every useful piece of information inside the company to have multiple opportunities to generate demand
15. Every month, we review what's actually generating revenue and double down on it
Which topics are attracting decision-makers?
Which lead magnets are generating qualified meetings?
Which executive accounts are reaching the right audiences?
Which content formats are producing the strongest commercial conversations?
We feed those insights back into the research engine, content strategy, lead magnets and sales process
Basically every LinkedIn GTM lesson we learn from one client becomes something we can test and improve across the rest of our clients
And the idea behind all of it is simple
LinkedIn is a complete distribution channel, an authority-building system, a source of buying signals and a way to generate qualified pipeline
So we connect every part of the process:
Research → Content → Distribution → Lead magnets → Qualification → Follow-up → Meetings → Revenue
And instead of hoping a few viral posts generate customers, we build an entire acquisition system around the people our clients actually want to close
That's how we've generated 50M+ views, 300,000+ followers and $5M+ in tracked revenue across our B
And it's the same system we continue improving every month
Todd Cohen retweeted
I came to the conclusion that most SaaS should become an MCP server.
we’re getting the economics of AI software backwards. A lot of AI startups are building their own chat interfaces, agents, and workflows, essentially reselling tokens from the big labs at low to negative gross margins.
What if SaaS companies did the opposite?
Take HR software. Instead of building an AI assistant, become the best MCP server for HR. Expose tools to:
- Benchmark salaries against proprietary market data
- Run compensation reviews and calculate pay equity
- Model salary increases against a department’s budget
- Recommend promotions based on company policies
- Update employee compensation in the HRIS
- Manage approvals, permissions, and audit trails
Let ChatGPT, Claude, or Copilot handle the conversation, reasoning, and orchestration. Let the customer pay for the tokens.
The SaaS company keeps the valuable stuff: proprietary data, integrations, business logic, permissions, compliance, and the system of record.
You still charge a SaaS subscription, but you’re no longer paying to generate every answer. You can ship faster, with fewer engineers, and potentially restore software-like gross margins.
The next generation of SaaS might not have much of a UI at all. Just an incredibly good MCP server.
Todd Cohen retweeted
This is how agents will ship all code.
→ Agent writes feature-gated code
→ Verification gates on push (proof checking, security audits)
→ Agent monitors metrics
You can live the future today with Vercel CLI: 𝚟𝚎𝚛𝚌𝚎𝚕 𝚏𝚕𝚊𝚐𝚜, 𝚟𝚎𝚛𝚌𝚎𝚕 𝚖𝚎𝚝𝚛𝚒𝚌𝚜, 𝚟𝚎𝚛𝚌𝚎𝚕 𝚍𝚎𝚙𝚕𝚘𝚢.
I haven't touched localhost in month.
I ask AI to make the feature only visible to me in production, so I can assess and iterate. If it's good, I release it to users, and ask the agent to set hourly/daily reminders to check logs and database to fix any issues. My AGENTS·md also has rules to force the agent to test everything.
Now AI is so good, everything goes straight from prompt to prod.
Here's my open source harness eval repo and what it found
github.com/garrytan/g-harnes…
Y Combinator 总裁 Garry Tan 开源 Agent 外壳评测库 g-harness-evals,实测同一大模型换不同框架通过率差不多,但单次成本差了 2.5 倍以上
Y Combinator 总裁兼 CEO Garry Tan 在 GitHub 上开源了一个名为 g-harness-evals 的评测项目。大家平时选编程 Agent 时往往只盯着底层大模型,但真正把聊天模型变成干活 Agent 的,其实是模型外面的那层“外壳”(Harness)——也就是系统提示词、工具调用、记忆机制和权限控制。
Garry Tan 搞这个项目的逻辑很直接:把底层大模型(如 Claude Opus 5.5、GPT-6.1 Sol)和测试任务完全固定死,只换外面的 Agent 框架,看看到底哪家干活最好、花钱最少。目前被拉上实验台对比的主流框架有五个:Claude Code、Codex、OpenClaw、Hermes Agent 和 QM,同时还顺带测试了他自己做的两个外挂组件:知识库 gbrain 和技能包 gstack。
【首轮实测:成功率拉不开差距,账单差距巨大】
第一批公开任务跑下来,有几个非常反直觉的发现:
在 20 个开发测试任务上,7 组“模型+外壳”搭配的任务通过率全都挤在 71% 到 79% 之间。比如同样搭载 Opus 模型,Hermes 通过率是 78.8%,Claude Code 和 OpenClaw 都是 76.3%,QM 是 72.5%。Garry Tan 在报告里直接写明,这些分数的 95% 置信区间高度重叠,统计上根本分不出高下。
真正拉开差距的是钱。每成功解决一个任务,OpenClaw 平均要花 0.44 美元,Claude Code 是 0.22 美元,而 Codex 只要 0.17 美元。
为了搞清楚钱到底花在哪了,团队把各个框架连到一个虚假的模型接口上,一分钱不花地抓包了它们发出的第一个请求。结果发现,面对一模一样的任务,不同框架塞给模型的初始上下文体量天差地别:Codex 大约发了 12,900 个 Token,Claude Code 发了 18,416 个,而 Hermes 直接飙到了 20,613 个。而且在一次 Sonnet 模型的测试里,OpenClaw 账单里 87% 的钱都花在了提示词缓存(Prompt Caching)的读写上,真正模型输出只占 13%。
【塞进上下文的技能包,模型一次都没执行】
Garry Tan 还拿自己写的两个热门开源组件开刀,结果一点面子都没给自己留。
第一个是 gstack(一套模拟工程团队角色的技能文件)。测试发现,虽然 gstack 的技能说明文本在每次会话里都成功加载进了 Opus 5.5,但在 46 次测试里,模型实际执行技能步骤的次数是 0 次,反而让单次测试成本直接翻倍。团队试着加了一个拦截钩子(Hook),强制要求先跑失败复现才准改代码,结果 Opus 直接通过 Bash 命令写源码,绕过了钩子整整 40 次。只有在加上专门的启动钩子(Start Hook)后,Agent 才会乖乖调用启动命令。
第二个是 gbrain(团队共享知识库)。如果不专门按提问者的身份做权限隔离(Scoping),直接给机器人一个全公司通用的 Token,在检索时 OpenClaw 和 Hermes 分别吐出了 123 条和 121 条提问者根本无权查看的跨权限私密记录。只有把知识库严格限制在提问者自己的权限范围内,越权读取才会降到 0。此外,挂载 gbrain 时,只有 Claude Code 会把知识库自带的操作说明传给模型,其他几个框架全都直接丢弃了。
整个仓库目前做成了“无密钥可验证”的模式。每一项跑分、置信区间和账单明细都附带了完整的原始记录(Receipts),任何人都可以在本地用 Bun 和 Python 3 跑脚本复算核对,不需要消耗自己的 API Key。Garry Tan 也在文档开头坦诚披露了利益相关:QM 是 YC 孵化的项目,YC 投资了 Hermes,而 gstack 和 gbrain 是他自己的个人项目。
Todd Cohen retweeted
Ok I've been thinking about this nonstop and I think it's bigger than most people realize.
Yet most business owners will NEVER set it up themselves.
I think there's a lot of money to be made here for AI consultants (just like there was in setting up Google Business profiles in the 2010s).
1. I'd lead with a free audit. Literally just ask ChatGPT about their business, open the profile, screenshot what's wrong or missing, then send it to them.
2. Sell the profile setup for $297-497. Claim it for them in Business Hub and walk them through verification. Then hand hold them through the Persona ID check on a 20-minute call (it needs their government ID so you can't do it for them).
3. Once you're in fill every field (description, category, logo, cover photo, website, delivery, return policies, etc). Retailers get a product feed with live prices and stock. Then have the owner add you as their agency. (never claim a client under your own account)
4. Upsell them on a Business Agent build for $1,500-3,000.
A Business Agent in ChatGPT is an agent attached to the business’s ChatGPT profile. It answers questions using ONLY what the business gives it.
-Clean up their product info and policies
-Write the agent's instructions
-Test 30-50 real customer questions pulled from support emails and reviews
-Fix every wrong answer
-Publish the agent
5. Sell the agent maintenance package ($300-500/mo). You help them add new products/price changes/policy updates. Every change gets updated and retested. Plus a monthly report from their profile insights.
This is exactly why the owner should be willing to pay for an agent instead of letting customers ask regular ChatGPT:
> Regular ChatGPT pulls from whatever it finds (reviews, forums, competitor comparisons). The agent answers from the info you give it.
> Regular ChatGPT lists you next to your competitors. The agent is a SEPARATE chat about your business (separate from their original ChatGPT thread).
> The product feed gives it LIVE prices and stock, not whatever it last crawled.
> It gives potential customers your return policy, your sizing guide, and your exact answer to "is this right for me?"
> It can collect contact details when the customer opts in (regular ChatGPT sends you nothing).
> It sends the customer to your site with their questions already answered.
> It's the landing spot for ChatGPT Ads once they're ready to spend.
I think this will be one of THE sexiest offers of 2027!
Now in beta: Businesses in ChatGPT.
Claim your business, manage your information, and get insights on how people engage with you.
Eligible businesses can also build a Business Agent to answer customers’ questions and help them explore what you offer.
openai.com/index/introducing…
Anthropic just shipped a 𝗯𝗶𝗴 𝘂𝗽𝗱𝗮𝘁𝗲 𝘁𝗼 𝗣𝗿𝗼𝗷𝗲𝗰𝘁𝘀 in Claude Code. I've been using it for a long time. It fully replaced running Claude Code on my own machine through Remote Control, and 𝗲𝘃𝗲𝗿𝘆𝘁𝗵𝗶𝗻𝗴 𝗜 𝗱𝗼 𝗶𝘀 𝗮 𝗽𝗿𝗼𝗷𝗲𝗰𝘁 𝗻𝗼𝘄.
→ Ask for a status update in the project chat and the coordinator tells you where each thread is
→ /btw: something else comes up mid-task, it gets its own thread and the one you're in stays on topic
→ /fork: want to try a different approach, it starts a new thread from the one you're in
→ Set a reminder on any thread, or just ask Claude to remind you. When it's time, it brings the thread up in the project chat
→ Share a single thread as a read-only link
It works on web, desktop and the mobile app. Projects is in public beta on Pro and Max. If you don't have it yet, join the waitlist.
Todd Cohen retweeted
higgsfield is robbing you
you shouldn't be paying $50 for a talking head AI avatar video with Seed Dance 2.5
So I made a open source workflow for you
You just give it the scripts, select the avatars you want to speak it, hit go, and it renders the videos
If creative velocity is something that you're trying to solve, this is a perfect way to do it
And then once these videos are generated you can upload them directly to the Facebook ads account via the Facebook ads API
get started free here - graphed.com/platform
give this skill file to your claude code - github.com/graphed-inc/marke…






























