Imbue

“The task that the biological brain has to solve is very, very different than what an artificial network has to do. To me, the clearest example of this distinction is, whatever solution the brain has learned to produce intelligent behavior has to go through this genetic bottleneck... But there’s no reason why artificial intelligence has to pass through a similar bottleneck.” In this Generally Intelligent episode, we spoke with Rylan Schaeffer, a Stanford PhD student studying the engineering, science, and mathematics of intelligence, about the importance of challenging dominant research ideas. Rylan authored the paper "Are Emergent Abilities of Large Language Models a Mirage?", as well as other interesting refutations in the field. He previously interned at Meta on the Llama team, and at Google DeepMind. Podcast links and highlights: https://lnkd.in/gYEUvKpn Papers mentioned in this episode: - Emergent Abilities of Large Language Models: https://lnkd.in/gE_jm6qi - Are Emergent Abilities of Large Language Models a Mirage? https://lnkd.in/epy4Yw-J - On the Stepwise Nature of Self-Supervised Learning https://lnkd.in/g8u4cH5a - The Curse of Recursion: Training on Generated Data Makes Models Forget https://lnkd.in/d9m4c_qt - Self-Consuming Generative Models Go MAD https://lnkd.in/d9P2XiAC

The latest Minus One episode dives deep with Imbue Co-founder and SPC alum Kanjun Qiu, exploring AI agents, the importance of play in human agency, the role communities serve in the tech ecosystem, and a lot more.

Why did we release our 70B model’s training process and toolkit rather than open-source the weights? Our CEO Kanjun Qiu explains the philosophy of empowerment behind our decision on Weights & Biases' Gradient Dissent podcast: "Imbue is all about directly empowering people. When I say 'empower,' it means a group of people feel like, 'okay, I know how to customize this to my needs. I know how to make this my own so that I can use it.' Our product, our culture, everything is around directly empowering people, and decentralizing and distributing power. Open-sourcing a model's weights is very different from open-sourcing the process for creating one. Even though creating a model might be expensive, [knowing] the process is much more empowering. If you read about the process and you understand it, and you apply it to some mid-training run, that's much more empowering than being given a black box." Read about our 70B model training process and access our toolkit here: https://lnkd.in/gpzE6wwM Thank you to Lukas Biewald for the great conversation! Links to the full episode: https://lnk.to/UtIbV2

🎙️ Generally Intelligent Episode 36: Ari Morcos on leveraging data to democratize model training We sat down with Ari Morcos, CEO of DatologyAI, to discuss: - how data washes out inductive bias - the “bitter lesson” of human-designed systems - the challenge of using synthetic data …and more! Podcast links and highlights: https://lnkd.in/gA-t2KFt

How can algorithms distribute agency and magnify curiosity? Glenn McDonald draws from his experience designing music algorithms at Spotify to examine how “systemically moral” algorithms can shift cultural validation from lotteries toward meritocracies. McDonald suggests that magnifying curiosity can be a means of decentralizing power: https://lnkd.in/g-h9Dc5m by Imbue

Great to see some really clear, practical, examples of how to use human judgement (sourced through Prolific) to finetune and improve model performance here: https://lnkd.in/eQZbJnDE

🤝 Imbue sourced vetted participants from Prolific to gain insight on question quality for their datasets. This data will help train their AI model on human-quality judgments. Check it out below or get started with Prolific for #AI here ▶️ https://lnkd.in/eCQKE3CE 🔗

Early this year, we trained a 70B model optimized for reasoning and coding. This model roughly matches LLAMA 3 70B despite being trained on 7x less data. Today, we’re releasing a toolkit to help others do the same. Read more and access the toolkit here:

Early this year, we trained a 70B model optimized for reasoning and coding. This model roughly matches LLAMA 3 70B despite being trained on 7x less data. Today, we’re releasing a toolkit to help others do the same, including: • 11 sanitized and extended NLP reasoning benchmarks including ARC, GSM8K, HellaSwag, and Social IQa • An original code-focused reasoning benchmark • A new dataset of 450,000 human judgments about ambiguity in NLP questions • A hyperparameter optimizer for scaling small experiments to a 70B run • Infrastructure scripts for bringing a cluster from bare metal to robust high-utilization training …and more! Read more and access the toolkit here: https://lnkd.in/gpzE6wwM Along with our tools, we’re sharing three blog posts with learnings from our training process: I. Conducting evaluations We found that our model and the best open-source models, when fine-tuned, outperform GPT-4o zero-shot across most multiple choice benchmarks. Surprisingly, both open and closed models achieve nearly 100% accuracy when evaluated only on unambiguous questions. We cleaned our evaluation datasets to isolate true failures of reasoning from failure due to ambiguous or low-quality questions. https://lnkd.in/giNS4z6h II. Setting up infrastructure Using our cluster for high performance training meant that every component — InfiniBand, Ethernet, GPUs, and the nodes themselves — had to work perfectly. If even a single one of the over 12,000 connections was a little flaky, it could slow down the entire training run. We're sharing open-source scripts and an end-to-end guide for infrastructure set-up that details the process of making everything work perfectly, and ensuring that it stays that way. https://lnkd.in/gCPDBknu III. Scaling experiments We successfully scaled from a 7B run to a 70B run on the first try, with minimal training instability and no loss spikes. We also predicted performance of the 70B model based on experiment results from much smaller models. We accomplished this using our hyperparameter optimizer, CARBS. We’re open-sourcing CARBS today so that other small teams experimenting with novel model architectures can experiment at small scale and trust performance at large scale. https://lnkd.in/gTgSjBvm This is one of many projects we’re working on to build collaborative agents that can reason and code. Other areas include RL, data generation, and experience design to make these powerful capabilities accessible and intuitive to users. We're hiring: https://imbue.com/careers/

"With AI, the default path is more centralization of power. This is why people are scared of the idea of 'AGI,' because AGI is power being centralized in a single entity. It takes a lot of effort and invention to buck that trend and actually make these systems democratize power. The real potential of agents, if we do a good job with invention, is that we can give individual people much more power over our computing environments." — Our CEO Kanjun Qiu at Collision Conf 2024, in conversation with Fast Company's Global Tech Editor Harry McCracken