Meta may face a $7 billion hit to its advertising revenue in 2025 due to escalating US-China trade tensions and tariffs. Analysts attribute this risk primarily to anticipated ad spend reductions from retailers like Temu and Shein, which account for a significant share of Meta's $18 billion in revenue from China. Although Meta has no user base in China, Chinese advertisers have relied heavily on Facebook and Instagram to reach global markets, making the country one of Meta's top revenue sources only after the US.
Perplexity plans to track user activity through its upcoming browser, Comet, to collect data for targeted ads. CEO Aravind Srinivas said work-related queries aren't enough to build detailed user profiles, so it wants data on shopping, travel, and browsing habits. The company expects users to accept this in exchange for more relevant ads and has partnered with Motorola to pre-install its app on Razr phones. Perplexity and OpenAI also expressed interest in buying Chrome if Google is forced to divest it amid an ongoing DOJ antitrust case.
Many B2B companies focus heavily on onboarding and renewals but neglect the long "middle" of the customer lifecycle, where engagement often drops off. This period is overlooked due to unclear metrics, ownership confusion, and the difficulty of creating content for established users. This leads to missed expansion opportunities, declining usage, and last-minute renewal scrambles. To address this, companies should implement milestone-triggered communications, create mid-lifecycle value roadmaps, and develop re-onboarding experiences to keep customers engaged.
AI search is reshaping how visibility works, and traditional SEO alone is no longer enough. The AI Visibility Pyramid introduces a layered strategy that starts with strong SEO fundamentals, moves into context-aware content that answers specific queries across AI platforms, and builds trust signals through third-party mentions and digital PR. Brands that consistently execute across these levels are more likely to appear in AI-generated results in the new search context.
To prevent unnoticed issues with their demo form, the team at MadKudu set up an AI browser agent that mimicked a real user and tested the booking flow daily. The agent fills out the form, waits for the calendar to load, selects a time, and flags anything that doesn't work, ensuring the full user experience is functional, not just the form. The team used Convergence.ai and set it up in 5 minutes by writing a plain-language prompt and scheduling daily tests. This setup now alerts them immediately if something breaks, helping the team avoid lost leads and surprise errors.
Snapchat has announced a new affiliate program designed to drive ad sales. Marketers, publishers, and creators can earn rewards for each qualified advertiser they refer to the platform.
CrewAI enables businesses to automate complex workflows by orchestrating multiple AI agents, referred to as "crews," which collaborate to perform tasks. Users can build these workflows using CrewAI's framework or UI Studio, choosing between coding from scratch or utilizing no-code tools and templates. The platform integrates third-party AI models from providers like OpenAI and Anthropic.
Anthropic's chief security officer has warned that AI-powered virtual employees could appear in corporate networks within a year. This raises urgent cybersecurity concerns. These AI identities would have autonomy, roles, memories, and corporate accounts, challenging current systems for access and accountability. Anthropic is testing its Claude models for resilience and misuse prevention while cybersecurity firms race to develop tools for managing non-human identities.
Effective brand protection is now much more focused on anticipating reputational risks and addressing them through strategic brand experiences than reactive PR. Audiences do not respond well to corporate statements as much as they respond to people especially before a crisis happens. Companies that empower internal teams to communicate externally, that show up in overlooked but high-impact environments, and that reinforce their brand promise through real-world experiences outperform competitors when trust is tested.
Reasons include slow page load speed, not having a FAQ section, and too many competing CTAs.
A table that maps key jobs to be done to the most effective AI tool.
The Everyone Hates Elon campaign illustrates a new model of guerrilla activism that blends meme culture, public installation, and social media to mount reputational challenges against high-profile individuals.
Sponsored
With Metabase Embedded Analytics [4], you can go from backlog to interactive dashboards in a matter of days, not sprints. Skip the never-ending DIY approach — Metabase handles filtering, drill-through, querying, auth, bug fixes, and all the feature creep you didn't plan for. Fully customizable to match your app's look and feel, with options from simple iframes to full React SDK.
Traditional methods of prioritization in product development - such as intuition, consensus, and rules of thumb - may be outdated, prompting a shift toward more structured frameworks like Cost of Delay. The focus has shifted to evidence-guided development, where prioritization is a continuous, evidence-informed process embedded within product discovery ultimately aimed at enhancing judgment rather than replacing it.
PMs can stay relevant with AI without coding by applying product thinking to real use cases, customer conversations, and internal workflows. Mapping AI concepts to domain problems and running lightweight experiments unlocks tangible product value.
2025.04.25·amycmitchell.substack.com·6 min read·TLDR Product Feature bloat silently erodes product value, costing time, money, and clarity. Teams can reclaim focus and boost engagement by using structured tools like the Feature Sunset Scorecard and communicating removals thoughtfully.
PMs often hesitate to make strong recommendations due to organizational politics, lack of confidence, or past discouragement. However, they can regain influence and credibility in decision-making by exploring options, presenting evidence, and stating clear preferences.
2025.04.25·oneknightinproduct.substack.com·4 min read·TLDR Product Sponsored
When consumers expect instant payments, slow money movement is a competitive liability. Join leaders from the Federal Reserve, J.P. Morgan, and others at Modern Treasury's Transfer 2025 [10] to explore the intricacies of building real-time payment infrastructure in the real world. Attend live in San Francisco, or join the livestream from
AI is transforming SaaS pricing by introducing marginal costs, pushing companies toward hybrid models like bundled usage, AI credits, or add-ons. Companies must balance cost control, user experience, and revenue growth while preparing for AI-driven changes in product value as usage affects profitability.
The quality of team conversations mirrors organizational health, highlighting the need to design environments that support meaningful dialogue. Effective communication relies not just on skills, but also on culture, structure, and systemic conditions. Leaders should ask which conversations are happening and why.
Effective user onboarding leverages behavioral psychology by simplifying complexity, celebrating progress, and visually guiding actions. Teams can design onboarding flows that boost user confidence and foster long-term engagement by applying principles like Miller's Law, the Goal Gradient Effect, and the IKEA Effect.
Startups that master marketing attribution gain a powerful edge: they know exactly which campaigns drive results. Teams can scale spending confidently, optimize faster, and build a predictable growth engine by implementing campaign-level, end-to-end attribution early.
Tech investor Keith Rabois shares his philosophy on embracing stress, spotting talent early, and valuing input-driven over output-driven cultures. He also discusses his contrarian views, influences like Margaret Thatcher, and his focus on discipline, resilience, and proactive approaches to modern challenges.
2025.04.25·thegeneralist.substack.com·12 min read·TLDR Product Maintaining product-market fit is increasingly challenging, especially in the AI era, where rapid growth can quickly be followed by decline.
Managers play a smaller role in promotions as professionals advance, requiring a shift towards self-advocacy, cross-functional collaboration, and broader influence.
Sponsored
If slow QA processes and flaky tests are a bottleneck for your engineering team, you need QA Wolf.
Devs can build ultra-fast websites by focusing on user journey metrics like server response, rendering, and interaction speed. Using Next.js features such as pre-rendering, caching, and React Server Components can help optimize performance. Each technique adds up to a smoother overall user experience over time.
This post discusses a class of React components that attempt to mix server-only and client-only logic, such as reading a file from the filesystem (server-side) while handling user input (client-side). Such components are “impossible” in traditional React rendering models. This tension illustrates the need for clear boundaries between server and client responsibilities.
DoorDash replaced Elasticsearch with a custom-built search engine using Apache Lucene to better handle complex relationships, custom ranking, and large-scale indexing needs. Its architecture separates indexing and search, supports fast segment replication, and includes a declarative schema system with query planning.
There's a difference between AI model and AI application innovation, and value will accrue differently in each area. Startups can innovate by focusing on specific use cases and integrating deeply with customer workflows, which happens in the AI application side of things.
There are four key criteria for vetting potential employers: fair pay, good treatment of employees, financial stability, and openness to new ideas. While a job may seem great initially with appealing perks, employees are ultimately a transaction and should not sacrifice their well-being or values for their employer.
node-mlx is a Node.js machine learning framework based on MLX. It has support for GPU acceleration on Macs with Apple Silicon and CPU support on x64 Macs and x64/arm64 Linux. node-mlx provides JavaScript APIs that mirror the official Python MLX APIs with some adjustments for JavaScript limitations. The repository includes examples for large language models, vector databases, and image search.
OpenAI has released gpt-image-1, a new image generation model based on the model that powers image generation in ChatGPT, in its API. This API allows devs to integrate high-quality image generation into their own tools and platforms with features like diverse styles, custom guidelines, and accurate text rendering.
This blog post goes over a high-performance and cost-efficient PDF rendering pipeline using Rust, AWS Lambda, and Typst. The solution uses AWS services like API Gateway, SQS, Lambda, and S3 along with performance optimizations like caching and batching. The resulting system had a rendering latency of 34ms and a cost of $0.35 per million PDFs. It was able to generate 1 million PDFs in 10 minutes.
Social media has evolved into a platform for content consumption, with Meta's Facebook and Instagram leading the shift. The FTC is challenging Meta's acquisitions of competitors like Instagram and WhatsApp, arguing they reduce competition, while Meta claims the market is too diverse for monopolies to exist.
Colanode is an open-source, self-hostable collaboration workspace with features like real-time chat, rich text pages, customizable databases, and file management.
The DOJ's antitrust case against Google has proposed forcing the company to sell Chrome - OpenAI has expressed interest in acquiring the browser.
A tool that packs your entire repository into a single, AI-friendly file.
COUNT(*) can be slow in PostgreSQL due to MVCC and the need to verify row visibility, but optimizations like filtering on indexed columns or using approximate counts from pg_class can speed the query up.
Sponsored
Discover the 7 essential concepts, techniques, and mitigation strategies for securing your AI pipelines.
Tesla is now testing its autonomous ride-hailing service with employees in the Austin and San Francisco Bay Area. The company has completed over 1,500 trips and 15,000 miles of driving. The robotaxi service will launch in Austin in June with 10 to 20 vehicles set to launch on day one. It will use Tesla's existing portfolio of vehicles rather than the Cybercab concept the company debuted last year. Autonomous vehicle companies in California need a range of permits to operate - Tesla so far only has a permit to test autonomously with a safe driver.
2025.04.24·techcrunch.com·3 min read·TLDR Google Gemini reached 350 million monthly active users as of March 2025, with a daily traffic of 35 million users. The number of users has increased significantly since late last year, when the service had tens of millions of users and 9 million daily active users. Despite the uptick in usage, Google is still far from catching up with OpenAI. The company is likely losing money on each user - OpenAI has previously noted that it loses money even on its $200 monthly plan.
2025.04.24·arstechnica.com·2 min read·TLDR Researchers have figured out a way to artificially expand the natural human color gamut. Humans usually have three different types of cone cells that combine their signals to form the normal color spectrum. The 'green' cones' function overlaps completely with the 'red' and 'blue' cones, meaning there's no wavelength of light that stimulates only the 'green' cones under natural conditions. Shooting a green laser at only the 'green' cones creates a color that the brain is unfamiliar with - the researchers have named the new color 'olo'.
2025.04.24·sciencealert.com·4 min read·TLDR A new study conducted by researchers at the Center for AI Safety, MIT's Medicine Lab, Brazilian university UFABC, and SecureBio claims that AI models can now outperform PhD-level virologists in problem-solving in wet labs, where scientists analyze chemicals and biological material. In the study, virologists scored an average of 22.1% in their declared area of expertise while OpenAI's o3 reached 43.8% accuracy and Google's Gemini 2.5 Pro scored 37.6%. The results are concerning to some scientists as for the first time in history, virtually anyone has access to a non-judgmental virology expert that could walk them through complex lab processes to create bioweapons.
You shouldn't just store the minimum data required to satisfy the current product spec, you should also store data that you'll highly likely use, such as timestamps and contextual metadata. For example, instead of just deleting rows from a database, engineers should instead store data on who deleted it, how they deleted it, when, and why. While not every bit of data stored will end up serving a purpose, the extra data could end up saving you one day, which makes up for the cost of storing it. While you shouldn't log just everything, people rarely complain about a table having too many timestamps.
2025.04.24·scottantipa.com·2 min read·TLDR OpenAI's new and improved image generation feature is now available on its API. Developers are able to generate multiple images at a time, control generation quality/speed, and adjust moderation sensitivity. All images will be watermarked with C2PA metadata so they can be identified as AI-generated by supported platforms and apps. It costs around 2 cents to generate a low-quality square image and 19 cents for a high-quality square image.
2025.04.24·techcrunch.com·2 min read·TLDR Claire Butler joined Figma in 2015 as the first marketer and business hire. Since then, the company has grown from nine people coding quietly in a small office attached to a rental its CEO shared with its co-founder to 1,400 employees and millions of users. Butler just left the company - this post contains a succinct list of lessons that Butler learned over the decade.
2025.04.24·clairebutler.substack.com·15 min read·TLDR