35,800 results · Most relevant
The best version of Reducto is one where you can abstract away the PDF processing problems, even if the PDFs contain sub-intricacies. Reducto's goal is that every single page that needs to be read by a model is read through Reducto first, and that this needs to be frictionless.
Adit Abraham argues that manual data entry from physical documents is the first category of work AI document understanding should automate, because it is rote, routine, and adds weeks of work. Humans have a lot to add in reasoning and decision-making, so they should retain those roles while AI handles transcription of physical documents. There is still a large amount of human-in-the-loop document work that can be automated today but has not yet been.
Adit Abraham recommends putting a ton of time into not just model architecture but also the data pipelines around it: evaluation infrastructure, RL loops around post-training/mid-training, and feedback loops from production errors — find mistakes, understand why they happened, and build better representations. Working with the hardest data distributions (largest enterprises, leading AI-native companies) helps you understand corner cases where different types of models fail.
For every magnitude of accuracy unlock, more use cases move into the fully automated range — use cases that were previously unimaginable become possible. Reducto has seen this repeatedly: when a new frontier (like accuracy) is unlocked, new applications appear that no one had imagined.
Reducto cares a lot about not just accuracy, but determinism and traceability — customers need to know where an answer came from. Traditional CV was really good at determinism and traceability, which is why it was used for structure preservation.
Focus is really important: the fact that Reducto can constrain the set of problems they need to be good at is a big part of why they're able to be ahead. A vision encoder for a model powering ChatGPT needs to be able to take a selfie and turn it into a Studio Ghibli image, whereas Reducto knows the domain of information they work around and needs their models to be exceptional on that specific set of tasks.
When your recommendation system starts failing, test a fundamentally different signal type (e.g., human critics) rather than tuning the existing algorithm. Maybe the best recommendation isn't a crazy complicated algorithm at all — maybe it's just humans, finding the critics whose taste is closest to yours.
Using shared credits (director, cast, genre) as a quality predictor leads to recommending bad films alongside good ones. If two films share director, cast, and genre, do not assume similar quality — many actors, writers, and producers have incredible and terrible movies in their portfolio. Common denominators are not necessarily the best quality or taste predictors.
Relying on similarity-based recommendations (same actors, same director, same genre) leads to watching effectively the same movie repeatedly — the speaker felt like he had watched the same movie five times in a row. If stuck in a recommendation loop, introduce randomness into your selection method rather than relying purely on similarity matching; the speaker gives Netflix credit for having some randomness that sometimes surprises him with very different stuff.
Chris and his crew follow the Entrepreneurial Operating System (EOS), a system that helps teams clarify and align their vision to execute plans. They use clear swim lanes between roles: 'I find the clients, you keep them.' Clear swim lanes keep you focused on different activities and are very effective — though they don't mean no communication: you still talk, discuss, and cross over, but by and large lanes hold.
Chris strongly believes you have to out-execute all of your competitors, because there is so much white noise everywhere that many people underestimate what it takes from an execution standpoint to really get results.
The expertise, authoritativeness, and trustworthiness component often comes from stories — not from externally crawlable things, so you have to create them yourself. Chris Dreyer creates E-A-T signals through manual interviews. Stories from directories and awards create better content that stands out from similar content across the web, and you can post to top directories and win awards as a way to generate story material.
People in the agency space struggle because they go too broad instead of achieving product-market fit.
Deep context — knowledge base, software standards, past decision archaeology, team methods, and judgment about who decides what — should be encoded into repos that both agents and humans can add to and correct, so it is built once and reused rather than rebuilt per mission.
Gary explains the structural change behind modern platforms: social media was originally coined on the premise that you'd get content from people you followed — your social circle. Interest media changed that: content is now served based on your behaviors on the platform, not just who you follow. He notes the algorithms and tech behind social platforms are changing, accelerated by AI over the last two years. He also cites Tumblr in 2008 as an early foreign concept — you followed art or music, not people.
Gary Vaynerchuk warns all creative people that demonizing quantity is the great vulnerability of their career. He argues quantity is objective, not subjective — you either make 800 pieces of content or you make one — while quality is subjective: production value, substance, lighting, and audio are all debatable variables. He observes that quantity is a very consistent variable in the brands and people that have elevated over the last three years, though he acknowledges you can also win on quality with one meaningful piece — both exist and both matter. His take: fall in love with 'and', not 'or'.
attributed_opinion
Work splits into two parts: the grind (reading, gathering, data cleaning) and judgment (the decision). Machines are incredible at the grind while humans are incredible at decision-making, so the two should be split into separate parts. With so much information now, the biggest question people have is what is the truth — what are the different sources saying, what is the divergence of information, and how can I trust what is out there.
endorsed as the speaker's proposal
Gary Vaynerchuk argues your content is posted and consumed in feed, and your audience is no longer the people who follow you — it's the people interested in the content you just made. If someone follows you for fashion and you post a marketing video, they'll just skip through it; but when his next post is about live shopping strategy to grow your personal brand, people stop. He notes that for a massive influencer, less than 5% of audience consumption happens on their grid, and the time people spend making their grid aesthetically pleasing — even though it hurts their views — blows his mind.
Gary Vaynerchuk states that the first strategic thing to understand is that every one of your companies is a media company that happens to be selling a product.
Gary Vaynerchuk frames quantity versus quality as one of the most interesting debates of the last decade, then takes a firm position: quantity is not debatable — you either make 800 pieces of content or you make one — quality is subjective, and quantity is a meaningful competitive advantage. He points to the receipts: among brands and people that have elevated over the last three years, quantity is a very consistent variable.
Choose sources → to get a library of your own and a daily brief.