AI models typically learn by memorizing patterns, then researchers bolt on reasoning as an afterthought. A new method called Reinforcement Pre-Training flips this approach—teaching models to think during basic training instead.
Good Morning from San Francisco, Meta's AI app turned users into accidental exhibitionists. 😬 People shared medical questions, legal
Meta users think they're chatting privately with AI. Instead, they're broadcasting medical questions, legal troubles, and relationship problems to the world through a public feed that many don't realize exists.
A new tool helps catch data mistakes before they mess up AI systems. Recce, which makes data review tools, just raised $4 million and launched a cloud platform to help teams spot problems early.
The startup tackles a growing problem: As companies build more AI systems, checking the data that feeds them becomes crucial. One wrong dataset can throw off an entire AI model.
"Most code reviews will become data reviews as data correctness becomes key for success in AI," says CL Kao, Recce's founder and CEO. Kao knows what he's talking about - he built SVK, an early version control system that companies like Apple and Ubisoft used before Git came along.
Recce's tools plug into existing data workflows and let teams compare datasets, check results, and track changes. Think of it as quality control for data pipelines. The company's open-source version already sees 3,600 weekly downloads on GitHub.
The Philadelphia Inquirer uses Recce to check data across 50+ systems. "It's helped us move faster without compromising data integrity," says Brian Waligorski, their lead data engineer.
Today's launch adds new features to catch problems earlier:
Heavybit, which led the funding round, sees this as perfect timing. "Data pipelines are the new secret sauce for companies building AI," says Jesse Robbins, Heavybit's general partner who joined Recce's board.
Other investors include Vertex Ventures US and Hive Ventures. Brian Behlendorf, who helped create Apache, also joined the board. He points out that AI adds randomness to software development, making data testing vital.
The money will help Recce build its cloud platform, which launches in private beta today. The platform lets teams:
This matters because AI systems are only as good as their data. Wrong data means wrong results, and catching errors early saves time and prevents problems.
Companies already use tools like dbt to handle data analytics. But as AI grows, they need better ways to check the data itself. Recce fills this gap by bringing standard software testing practices to data workflows.
The timing fits a broader trend. As AI becomes standard, unique datasets give companies an edge. But that only works if the data stays accurate and clean.
Kao will speak about these challenges at the Data Council conference in Oakland on April 24. The company also opened applications for early access to its cloud platform.
Why this matters:
Read on, my dear:
Three ex-Cisco executives raised $21M to build AI that handles system outages before engineers get woken up at 2 AM. Their secret: multi-agent systems that cut response time by 90%. But there's a catch that separates them from every other AI monitoring startup.
Perplexity AI, known for providing direct answers instead of traditional search links, is nearing a $14 billion valuation. Yet beneath the soaring numbers lies a pressing question: Can its novel approach truly reshape how we search the internet?
Bugs just got more expensive. Lightrun, an Israeli startup, has secured $70 million to tackle a peculiar side effect of
Benchmark just placed a massive bet on Chinese AI startup Manus, leading a $75 million funding round that values the company at half a billion dollars. Not bad for a startup whose AI agents sometimes crash while ordering takeout.
Get tomorrow's intel today. Join the clever kids' club. Free forever.
