Sara Saab

Just over four years ago, I met a YC-backed team in London working on an academic research marketplace. They charmed my socks off and seemed so hungry to take product, tech, and culture to new levels. They barely knew it then, and I _certainly_ didn't, but they would go on to close the UK's second largest Series A of 2023, become one of Europe's most important AI ecosystem companies, and be regarded as one of the world's top ethical human data companies for AI, industry, and academia. Human feedback at the speed of APIs. Literal moments from first touch to an opinion or annotation from a guaranteed-real human anywhere in the world. Unshakeable principles around fair rewards and fair play. We're talking autoresearch before autoresearch was a word. We're talking squishy humans in the loop without breaking automations and orchestrations. That company is Prolific. After 4 indescribable years of laughter, toil, mistakes, and ultimately, triumphs, I leave the Product, User Research, and Design team at Prolific in some of the best hands humanly imaginable. I know these amazing leaders (tbh every single hire at Prolific is a leader in some way!) are about to smash through some new ceilings on the way to the moon 🛫 🌕 What's next? Assuming Prolificos go easy on me at the leaving party and I survive the night 😁😁... I am INSANELY excited for the Next Big Thing (!) I'll say more soon, but as a teaser: we don't get to coast into the future of the internet and digital data with the same rules and norms that got us this far. Not only does the internet as a system not sustain it, but neither does human society at large. At some point over the last 2 years, we passed through a gate, and we have some pretty damn urgent work to do to intentionally build the right future. I will be doing some of that work, alongside some absolutely amazing thinkers and builders. But for now, here are some photos of silliness cos silliness is a big part of the Prolific DNA! Too many people to tag, but Prolificos: thank you all, I'm forever cheering you on, and see you down the road team. #keepthelidonthefort Sara Team Prolific 💙

Today Anthropic announces it’s watermarking Claude’s generated text. Apart from my obvious support from a social good perspective, I am really interested in the mechanics and the richness / ruggedness / durability / gameability of text watermarking. Like others I assume this uses a probabilistic algorithm similar to that described in Dathathri, See, Ghaisas et al (2024). As such, and unlike image watermarking, this text watermarking method is unlikely to carry a richer provenance payload. I also wonder how easy it is to break or reverse engineer and how much of a cat and mouse this turns into over time. The SynthID folks gave a lot of thought to how open to be with detector access to prevent adversarial exploration of the system, but ultimately seemed confident enough to (plan to) open it up. Curious what Anthropic does here! Intrigued / curious and glad teams are taking a swing at something that seems wickedly hard at first glance (hiding a clue to origins in text! It’s like a murdle puzzle!). Hope the team publishes a bit of science around this so we’ll be able to see for ourselves :)

Yet another Leonis Capital banger. I've been musing for a year or so that model swaps (and the evals undergirding them) are gonna be a huge economy, this is a much more cogent analysis of the dynamics behind it. Excerpt: "The next phase is therefore not less enterprise AI adoption, but a shift from access to allocation. In the first phase, companies bought seats, encouraged experimentation, and treated token consumption as a proxy for adoption. In the next, they will manage AI as a portfolio: centrally governed multi-model gateways, budgets assigned by workflow and risk level, cheaper defaults for efficiency work, and flexible frontier spending for expansion work or high-stakes judgment. We expect model access will become an escalation policy rather than a company-wide entitlement, and procurement will increasingly measure cost per accepted task and value per dollar instead of seats, benchmark scores, or raw token volume." lnkd.in/erFcpqcU