Kyle Waters

@r0ck3t23 But RL environments are still data... bottleneck has just shifted to designing rubrics & evals for harder-to-verify tasks in knowledge work (law/finance) & frontier scientific research. You still need experts in the loop to define objective criteria for rewarding success.

Dario Amodei just revealed that the AI training bottleneck everyone is worried about doesn’t exist anymore. The industry spent years obsessed with scraping the open web. More data. More text. More human output to feed the models. Amodei: “I don’t think data is quite the most central thing anymore.” The shift is fundamental. Amodei: “Static data is becoming less important. A lot of the data we use today is RL environments that we train on. Dynamic data that the model creates itself.” Not scraped. Not licensed. Not written by humans. Generated by the model through pure trial and error. When you train on complex math or agentic coding, you don’t feed it a textbook. You give it an environment. The model experiments. Fails. Adjusts. Tries again. Amodei: “You’re getting some math problems and the model experiments with trying the math problems.” It generates its own experience. Millions of iterations. Each one building on the last. No human required. This destroys the entire narrative around AI hitting a data wall. You cannot throttle a competitor by locking down copyright. Cannot slow the race by putting up a paywall. When a model learns through its own synthetic experience, the open web becomes irrelevant. The only true bottleneck left is compute. And this is where the geopolitical stakes become impossible to overstate. The nation that wins the compute race doesn’t just build smarter models. It builds models that generate their own intelligence, compounding on themselves, iterating past every limit human knowledge ever imposed. We are no longer training AI on the past. We are letting it simulate the future. The machine has stopped reading the dictionary. It’s doing the math itself now.

Amazing couple days of conversations at the @PyTorch Conference in SF - very clear that evals & data sourced from subject-matter experts will play a critical role in advancing AI performance in economically valuable settings.

IMO — Ilya is wrong - Frontier LLMs are are trained on ~200 TBs of text - There's ~200 Zettabytes of data out there - That's about 1 billion times more data - It doubles every 2 years The problem is the data is private. Can't scrape it. The problem is not data scarcity, it's data access. The solution is attribution-based control (article below) "Unlocking a Million Times More Data For AI"

8/ The lack of any reliable data valuation framework is a massive blocker to surfacing novel datasets for AI. Auctions are remarkable engines for pricing non-alike goods. The data economy deserves the same foundation, and it’s worth building.

7/ We've also started exploring a data valuation framework with some of our early users on the Datalab. We're still refining it, but it takes into account a dataset's key features like uniqueness, quality, modality, freshness etc.

1/ AI isn't just a compute race anymore. It's a data race too. Labs are paying top dollar for differentiated, high-signal data. It's clear now is the time to experiment with new approaches to valuing and incentivizing the creation of frontier AI data. x.com/LucasNuzzi/sta…

Noticing a trend? Specialized models continue to beat foundation models on task performance, cost, and latency. The emerging design pattern for agents is a foundation-model-brain that can invoke the most optimal tool for a given task.

GPT-4b micro is a model trained exclusively on specialized biological data. It was used to reverse cellular aging with a 50x improvement in efficiency relative to previous approaches. A testament to the power of narrow AI + specialized data. Amazing overview by @rowancheung:

AI is selling off because we've reached a plateau of what foundation models can do without specialized tools. It's like we have an amazing operating system but very few apps running on it. The way forward is fine-tuning tools with better data: the next major trend in AI ⬇️

11/ Full post below! research.portexai.com/what-hugging-f…

10/ Hugging Face proves that a collaborative approach to creating novel datasets is propelling AI forward. We love what Hugging Face has built and hope to become its counterpart for monetized datasets with the Portex Datalab: datalab.portexai.com/explore

1/ New @PortexAI research studying the massive growth and reach of @huggingface, which is unquestionably the heartbeat of open source AI today. Its repository of datasets and models point to a bright future for fine-tuning, and also a coming market for proprietary datasets 🧵