If your company's AI bill doubled every 45 days, could anyone show what got twice as good? At 8090, token costs are growing at roughly that rate while the added productivity is estimated at only 5% to 10%. Most companies will not notice the gap until it shows up in their margins. Track AI cost and output by task, then send routine work to cheaper models and reserve the expensive ones for jobs where the answer actually improves. You should also know where prompts and employee corrections go. Those corrections can reveal how your company makes decisions to the model provider.
Your activation rate can improve while the company gets worse. AI can polish your ads and onboarding enough to bring in more signups, while the share of those customers still paying two months later falls. Most activation dashboards will show the first number and miss the second. A better funnel starts when someone first interacts with the product, then follows them through setup, aha, activation, and habit. Also track how much you spent for every account that survives month two. If signup conversion rises while month-two survival falls, you did not fix onboarding. You got better at enrolling the wrong people.
The Meta MCP is now available to developers. It allows developers to speak to Meta's platform in natural language to create and edit campaigns, ad sets, and ads, perform detailed reporting, manage catalogs and feeds, run A/B tests, and review activity logs. The platform was designed for advertisers using AI assistants and developers building AI-native ad flows. It provides structured, permission-scoped access to ads capabilities without the need for individual endpoints.
Qwen3.8 is now available on Alibaba's Token Plan, Qoder, and QoderWork. The model has 2.4T parameters and is continuously evolving. It is comparable to leading frontier models, with Alibaba claiming it is second only to Fable 5. The model will be going open-weight soon.
Anyone can design now. Give AI a screenshot and a prompt and it can produce something clean in minutes. However, "looks good" is not the same as knowing why it works, what feels off, or what to change next. Train that judgment by copying great interfaces, studying their spacing, and presenting work before it feels finished. Perhaps for your own team, ask everyone to rebuild one interface without giving the screenshot to AI. Learn to notice why one margin is 16 pixels, another is 24, and why changing either one makes the whole screen feel wrong.
Most managers tell someone to "own" a task, then get annoyed when the result is not what they pictured. This guide says to agree on four things first. What counts as done, what good work looks like, when it is due, and which resources the person can use. Then let them decide where, when, and how to do it. Those four questions separate delegation from abandonment. If the work comes back wrong, you can see whether the expectation was unclear or the execution failed.
Raising money isn't free - every funding round means selling a share of your company's ownership for money.
The best you can do while copying your competitors is deliver mediocrity.
Neil Rimer from Index Ventures warns of an imminent redistribution of AI-generated wealth, hoping it will be voluntary.
The Bay Area pulls almost 2.5x the capital of the entire New York ecosystem across every category combined. Value concentration in AI and B2B is more concentrated now than it was five years ago. Two metros combined take 72% of B2B dollars. Talent and capital didn't disperse - the narrative got ahead of the numbers.
Agents at Replit now investigate production incidents, review pull requests, answer questions, analyze business data, triage support tickets, research sales accounts, and improve the systems that power the Replit Agent itself. The company's expanding system of agents takes goals from people, gathers context, performs work, checks the results, and escalates when human judgment is needed. People still decide on which problems matter, make difficult tradeoffs, exercise taste, and take responsibility for the outcome, but they increasingly do not perform every step required to get there.
Companies selling data to frontier AI labs know the cash is real, but the contracts are not forever. There are six categories of products being sold to frontier labs: hours, judgment, worlds, verdicts, bodies, and rights. Each generation of models graduates past the data that trained it, so labs need fresh data to keep pushing the frontier.
The point of the onboarding process is to teach users how to use your product. The MCP onboarding process should give agents a step-by-step guide on how to walk users through their first time using your product. Agent-native support can be better than founder support. The field of agent-first interfaces is still in a nascent stage, so a lot can still change.
Cerebras' employees ask the company's internal knowledge base more than 15,000 questions every day. It has become one of the company's most widely adopted internal tools since launching 3 months ago. The knowledge base is used by humans, automations, and agents. This post details how Cerebras Knowledge was built.
Someone can discover you in ChatGPT, Google your company a week later, and show up in GA4 as Direct traffic. That is why AI search can look tiny in your analytics even when customers are using it. Only about 1% of people click citations in Google AI Overviews, and 70.6% of AI-referred visits are recorded as Direct. Before deciding AI search is doing nothing, add ChatGPT, Claude, Perplexity, and Google AI to your "How did you hear about us?" question. Then tag the same answers in your sales calls.
1Password for Claude is built on a zero-exposure architecture that only grants access at runtime. Claude never sees the vault item, password, or one-time code, and access is scoped to the current task and ends when the task is complete. Access requires user-consented biometric approval, and credentials are injected directly into the page. 1Password for Claude is now available for Mac for business, family, and individual plans.
Muse Spark 1.1 is now available on OpenRouter for developers in the US. Muse Spark is a multimodal reasoning model from Meta. Built for agentic tasks, it accepts text, images, video, audio, and PDF documents and returns text. The model has a 1-million-token context window. It is designed to orchestrate multi-agent workflows and supports structured output, parallel function calling, built-in search with citations, and configurable reasoning effort.
Natural language is a far more useful form of input than the indirect signals the industry had to live with over the past 20 years. It allows companies to read exactly what users want to accomplish and whether or not they're satisfied. Natural language inputs, paired with other signals, can help companies see the relationship between the performance of agents and revenue. Agent observability is not only an engineering concern - CEOs, product leaders, and customer-facing teams should be reading this data too.
Venture firms can no longer rely on IPOs, M&A, or secondary buyers to convert portfolio markups into DPI. Healthy private companies may need cash-flow-driven buybacks and capital return programs, supported by duration-matched credit, to create liquidity without sacrificing growth.
It's an open secret that AI agents aren't reliable because they're stochastic, prone to lying, expensive, and make mistakes. Businesses are responding by selling better cages, but AI is not deterministic. Nobody survived the cloud wars by having a better shovel - they survived by building moats. These companies are competing with model improvements, and the models will eventually win.
Open source AI is funded by companies shipping a deliberately unremarkable, general-purpose base cheaply and letting customers build value through customization.
If a customer says "interesting" after your demo, they are probably about to ghost you.
Product executives, especially those skilled in building with modern tools, are now commanding salaries akin to pro-athletes due to high demand and scarcity.
Investors, LLMs, and probably your friends all say "taste" is the only thing left to defend now that AI can write the code.
An electrician built a $12.99 load calculator that can replace a $500 service call. A plumber says one afternoon with OpenClaw got him further than a $40,000 consulting project. The interesting shift is the minimum market size. A tool this narrow never justified a developer's salary, but when the domain expert builds it himself, a few thousand electricians is a viable market. Pricing is the easy part: the $500 service call already exists, so $12.99 reads as a discount on money the customer was spending anyway.
A badly managed AI agent can cost more than the employee it was meant to replace. The money vanishes into loops: agents spending tokens redoing work nobody clearly defined. This article warns founders against giving AI agents vague jobs and assuming cheap tokens mean cheap labor. A badly managed agent can cost more than the employee it was meant to replace. It keeps retrying, replanning, and correcting work nobody properly defined, and every loop costs money.
Founders benefit from setting initial valuations below market rates, enabling easier future fundraising and sustainable company growth. This strategy avoids the pitfalls of inflated valuations and offers opportunities to sell stock later at a higher price. The approach aligns with maintaining disciplined capital management, preventing overcapitalization, and ensuring long-term success.
Everyone seems to be reaching $1 million ARR in record time. PostHog's account of its first 1,000 users is refreshing because it is mostly repetition. Hacker News worked, so the founders kept posting there. People liked candid stories, so they wrote about the launch, the fundraise, and six ideas that failed. No secret growth loop.
Zapier MCP connects Claude, ChatGPT, Cursor, and other AI tools to Gmail, Slack, Salesforce, and over 9,000 other apps. It can be set up in minutes with no terminal, config files, or coding required. Zapier MCP is included in existing Zapier plans. Guides on how to use the MCP are available.
The Domain SDK makes it easy to add, verify, monitor, and remove customer domains with one TypeScript API. Users just add a hostname to the platform they already run, and it returns the exact DNS records customers need, tracks them until they are ready, and removes them safely. The SDK features adapters for Vercel, Cloudflare for SaaS, Railway, Render, and Netlify. The tool is server-side only and requires Note 20+ or Bun.
Notion now supports opening Markdown files as a read-only preview. Users can then import them as Notion pages. A quick video demo of the feature is available in the thread.
Startups have structural advantages over big model companies. Entrepreneurs should focus on leveraging these strengths rather than worrying about competition. This approach offers more opportunities for innovation and agility.
One of the most valuable life skills is seeing people clearly. Seeing people clearly allows you to better appreciate them, help them grow, and put them in a position to succeed. The best references are from those who see you clearly.
Math and science olympiads produce exceptional founders by developing rigorous problem-solving, resilience under pressure, and communities that reinforce ambitious career paths.
False product-market fit can sometimes be even worse than no PMF, because you're ramping up activity and spending while running in a straight line towards a cliff.
Companies push users towards apps only so they can better harvest information from them.
Sponsored
Ads get ignored on social media but not in TLDR! Reach developers, PMs, marketers, founders and other tech leaders where they actually
The complexity tax of owning your own weights is too high for most tasks. Companies need a place where they can bring their workload and get help building a task-specific model on their data, tuned to their tasks, running on infrastructure they control. This would give them the control benefits of owning their weights without having the human complexity cost and tax. They would not just own a model, but the engine to continually build and fine-tune the weights.
The AI boom has produced a new kind of company with enormous capital and power and genuine technical brilliance. However, they have no durable business model at their own layer. Foundation models are inventory, and every release obsoletes the last one. Enterprises now buy intelligence the same way they buy electricity. The model is, by design, the replaceable part of the workflow.
The biggest winners of AI are businesses nobody would ever call AI companies. AI transformation creates value through revenue, cost, and risk. For low-margin businesses, the biggest lever is cost - even small reductions in operating expenses can create an outsized increase in earnings. The lowest-margin businesses can finally attack the coordination costs that have kept them structurally low-margin for decades. The next wave of AI winners will put agents behind the workflows of low-margin businesses and let the savings show up quietly in the operating model.
The best software engineers are dramatically more productive than their peers. This has caused every tech company to hunt for those rare individuals capable of generating extraordinary results. Sierra found that its engineers were getting 5x more done on some tasks by running agents in parallel with git worktrees. It then set up a six-person AI acceleration team to make everyone at the company more productive. This article explains what the company built and what it learned in the process.
Revenue is the usual evidence for product-market fit, and it is the one metric a determined founder can manufacture without fit by personally forcing every deal across the line. The B2B Standard replaces it with six behaviors that cannot be forced, including buyers pulling in 60% of first meetings, a 50% close rate on those pulled deals, and 95% of new customers reaching the leading indicator of retention within a month. Put the six numbers on one page and review them weekly. The exact cutoffs are debatable, but whether buyers are pulling or the founder is pushing is not.
2026.07.13·thephysicsofstartups.substack.com·7 min read·TLDR Founders Agent harnesses help agents take on longer-running, more complex work. Harness design is critical to token efficiency. This post looks at how Glean moved its harness to orchestrate tools for every query using code. The approach proved more efficient, even for simple queries, and reduced token usage by 24%.
Genkit is an open-source framework for building full-stack AI-powered and agentic applications for any platform. It features a conversational support assistant that remembers tickets and a copilot that works across several turns. Genkit's Agents API packages message history, the tool loop, streaming, persistence, and a frontend protocol behind a single interface. It is now in preview for TypeScript and Go.
Anthropic has created a new feature that helps users reflect on and refine how they use Claude. It lets users easily track and visualize how they use the AI tool. The reflection dashboard can be found in Settings on Claude for the web or the desktop app. It is currently available in beta for Free, Pro, and Max users who have memory turned on.
AI tools can learn how a company makes decisions from employee corrections, even when its source files stay private. A claims adjuster explains why a case deserves an exception; a sales manager flags an account as risky and says why. Satya Nadella calls this the Reverse Information Paradox. The company pays for the model and also supplies the judgment that improves it. Enterprise contracts now need to cover prompts, corrections, evals, traces, memory, and tuned weights, and the customer should be able to switch models without losing what the system learned. Otherwise, the model gets smarter, and the customer gets more locked in.
Start a company, or join an AI company that's already winning? The old tradeoff was stability versus ownership, but now the big labs are where the risky technical work happens, and a two-person startup can end up building a minor feature around their models. A spreadsheet can compare salary, equity, and outcomes. It can't say how much responsibility you'll get, what you'll learn, or whether the problem matters, and those questions beat how ambitious the choice looks from outside.
Mercor posted the seed deck it used in September 2023.