Blog

Every mngr plugin, explained

Fine-grained HTTP permissions for AI agents

How AI code review can make correct code worse

Blueprint: One-shot bigger coding tasks

There Will Be a Scientific Theory of Deep Learning

Keystone-Eval: Our agentic benchmark for self-configuring agents

How to usefully run 1,000 agents in parallel

Bouncer: Heal your feed

How Offload Works: Inside the Rust CLI that sped up our tests by 6x

A case study in testing with 100+ of Claude agents in parallel

Usefully run hundreds of Claudes in parallel with mngr

Introducing Keystone: Building self-configuring agents

Offload: We made agents run our test suite 6x faster

Vet: Prevent coding agents from making mistakes

Latchkey: Use plain old REST instead of MCP

LLM-based Evolution as a Universal Optimizer

Beating ARC-AGI-2 with Code Evolution

A more radical Imbue

Choices and Knives

Digital Freedom Depends on Access Rights: Matt Boulos in Lawfare

How we made sandboxed coding agents 10x faster to start

A Healthy Ecosystem for AI Agents: Introducing the AFI

Closed vs. open AI ecosystems: Kanjun on Decoder

Sculptor: the missing UI for parallel coding agents

How to build robust software systems with AI: Josh at the AI Engineer World's Fair 2025

Beyond the Prototype: Using AI to Write High-Quality Code

Empowering humans in the age of AI

Sculptor: Catch and fix issues as you code

One-person enterprise: Kanjun at the World Economic Forum Annual Meeting

The future of AI is human-centered: Kanjun on the Pioneers of AI podcast

Glenn McDonald: On Dungeons & Dragons, data alchemy, and radically open software

Keeping humans in the loop: Josh at the Democratize Intelligence Summit

Human agency vs. “agentic” AI: Kanjun on the NonZero podcast

A more empowering view of general agents: Kanjun on NYSE TV

Transforming AI agents into personal collaborators: Kanjun on the NVIDIA AI Podcast

The ethical and practical future of AI: Matt Boulos at the MIT Club of Northern California

The AI paradox: Kanjun at the 2024 Progress Conference

Rylan Schaeffer, Stanford: On investigating emergent abilities and challenging dominant research ideas

Agency in the age of AI: Kanjun on the Minus One podcast

“The AI startup Imbue is building a service that will let you DIY a bot for any task, large or small,” said Kanjun Qiu, its chief executive and co-founder. Rather than relying on companies to build bots for us, Qiu said Imbue's service aims to give people more control in a bot-powered world.

Reinventing AI agents: Kanjun on the Gradient Dissent podcast

These Are the 10 AI Startups to Watch in 2024

Ari Morcos, DatologyAI: On leveraging data to democratize model training

Kanjun Qiu, CEO and founder of the AI startup Imbue, which is working on agents that can reason and code, likens the state of agents to where self-driving cars were just over a decade ago.

From bare metal to a 70B model: infrastructure set-up and scripts

Training >70B LLMs on 10,000 H100 clusters: Josh on the Latent Space podcast

Open-sourcing CARBS: how we used our hyperparameter optimizer to scale up to a 70B-parameter language model

Training a 70B model from scratch: open-source tools, evaluation datasets, and learnings

Sanitized open-source datasets for natural language and code understanding: how we evaluated our 70B model

“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." — Kanjun Qiu, CEO

“The lack of openness has definitely been noticeable over the past year and it's put a damper on academic investigations,” — Josh Albrecht, CTO

Percy Liang, Stanford: On the paradigm shift and societal effects of foundation models

'We need legal protections from AI’s risks': Matt Boulos in The Hill

“I think of it as really being able to [help us] control information environments … once we have control over information environments, we’ll feel a lot more empowered,” Qiu said. “Every single person’s vision can come to life.”

“Technology that we build reflects the values of its creators,” says Qiu. “There’s a much more human way to think about building out this technology, in a way that’s empowering to people.”

Seth Lazar, Australian National University: On legitimate power, moral nuance, and the political philosophy of AI

“What I’m most excited about is that I think [AI] is one of the biggest things that can unlock our potential.” — Kanjun Qiu, CEO

Imbue joins Department of Commerce’s AI Safety Institute Consortium

Tri Dao, Stanford: On FlashAttention and sparsity, quantization, and efficient inference

The future of AI is human-centered

Imbue and Dell designed the custom system... to support rapid experimentation in developing AI systems

We've partnered with Dell to build a cluster of 10,000 Nvidia H100 GPUs to train our models

Developing agents that can reason and code: Kanjun and Josh on the No Priors podcast

Methodology for analyzing the individual comments to the NTIA’s AI RFC

The need for trust: Analyzing the individual comments to the NTIA’s AI RFC

“We want to enable agents that people can actually use and that work reliably, and ultimately the vision is to reinvent the personal computer so that it’s actually truly personal.” Kanjun Qiu

Imbue Raises Fresh Capital, Amazon's Alexa Fund Joins Investor Lineup

Why AI agents don’t work (yet): Kanjun on the Latent Space podcast

Imbue raises $200M to build AI systems that can reason and code

Imbue raises $200 Million Series B

Our technical hiring process

Bryden Fogelman: On coding agents and finding meaningful work at Imbue

In-person culture and habits

Where are the good policy ideas? Analyzing the NTIA's AI RFC

“The real challenge is building systems with robust reasoning,” said Qiu.

What's next in AI?: Kanjun and Reid Hoffman at Config 2023

On the Stepwise Nature of Self-Supervised Learning

Jamie Simon, UC Berkeley: On theoretical principles for how neural networks learn and generalize

Scaling Laws For Every Hyperparameter Via Cost-Aware HPO

Our submission to the NTIA's AI accountability RFC

Towards guardrails, not guidelines: a policy framework for powerful AI systems

Kanjun Qiu, the CEO of Generally Intelligent, an AI research company, has been buying advanced graphic chips since last year for her own servers, allowing her to ride out the current shortage.

“The biggest thing…that I find them to be helpful for is inspiring me as the user and helping me learn things that I wouldn’t have necessarily thought of on my own,” Josh Albrecht, CTO of Generally Intelligent

Update: The state of LLM ethical decision-making

Bill Thompson, UC Berkeley: On how cultural evolution shapes knowledge acquisition

“Over the next few months and years, we can expect much of the internet to get connected to large language models.” — Kanjun Qiu, CEO

Instead of relying exclusively on outside providers, the AI startup Generally Intelligent has begun purchasing its own graphics chips to power its research.

Ben Eysenbach, CMU: On designing simpler and more principled RL algorithms

Jim Fan, NVIDIA: On foundation models for embodied agents, scaling data, and why prompt engineering will become irrelevant

Sergey Levine, UC Berkeley: On the bottlenecks to generalization, why simulation is doomed to succeed, and picking good research problems

The state of LLM ethical decision-making

Noam Brown, FAIR: On achieving human-level performance in poker and Diplomacy, and the power of spending compute at inference time

Sugandha Sharma, MIT: On biologically inspired neural architectures, how memories can be implemented, and control theory

Nicklas Hansen, UCSD: On long-horizon planning and why algorithms don't drive research progress

Jack Parker-Holder, DeepMind: On open-endedness, evolving agents and environments, online adaptation, and offline learning

“But it really begs the question: what is understanding?” — Kanjun Qiu, CEO

Celeste Kidd, UC Berkeley: On attention and curiosity, how we form beliefs, and where certainty comes from

Archit Sharma, Stanford: On unsupervised and autonomous reinforcement learning

Chelsea Finn, Stanford: On the biggest bottlenecks in robotics and reinforcement learning

Introducing Generally Intelligent

Our approach to safety

Avalon: A Benchmark for RL Generalization Using Procedurally Generated Worlds (NeurIPS Spotlight Paper)

Among those contributing to the company's $20 million in initial funding (plus over $100 million in options) is Tom Brown, former engineering lead for OpenAI's GPT-3; former OpenAI robotics lead Jonas Schneider; Dropbox co-founders Drew Houston and Arash Ferdowsi; and the Astera Institute.

Hattie Zhou, Mila: On supermasks, iterative learning, and fortuitous forgetting

Minqi Jiang, UCL: On environment and curriculum design for general RL agents

Oleh Rybkin, UPenn: On exploration and planning with world models

VICReg: Tutorial and lightweight PyTorch implementation

A PyTorch implementation of SIMONe

Andrew Lampinen, DeepMind: On symbolic behavior, mental time travel, and insights from psychology

Yilun Du, MIT: On energy-based models, implicit functions, and modularity

Martin Arjovsky, INRIA: On benchmarks for robustness and geometric information theory

Jupyter Ascending: Sync notebooks to your editor

Yash Sharma, MPI-IS: On generalizability, causality, and disentanglement

Jonathan Frankle, MIT: On the lottery ticket hypothesis and the science of deep learning

Jacob Steinhardt, UC Berkeley: On machine learning safety, alignment and measurement

Vincent Sitzmann, MIT: On neural scene representations for computer vision and more general AI

Dylan Hadfield-Menell, UC Berkeley/MIT: On the value alignment problem in AI

Drew Linsley, Brown: On inductive biases for vision and generalization

Giancarlo Kerg, Mila: On approaching deep learning from mathematical foundations

Yujia Huang, Caltech: On neuro-inspired generative models

A PyTorch implementation of slot attention

Julian Chibane, MPI-INF: On 3D reconstruction using implicit functions

Katja Schwarz, MPI-IS: On GANs, implicit functions, and 3D scene understanding

Joel Lehman, OpenAI: On evolving intelligence, open-endedness, and reinforcement learning

A new AI research lab: Josh on the Machine Learning Engineered podcast

Cinjon Resnick, NYU: On activity and scene understanding

Sarah Jane Hong, Latent Space: On neural rendering & research process

Kelvin Guu, Google AI: On language models & overlooked problems

Generally Intelligent #0: A podcast for deep learning researchers

Understanding self-supervised and contrastive learning with "Bootstrap Your Own Latent" (BYOL)

Appendix for "Understanding self-supervised and contrastive learning with 'Bootstrap Your Own Latent' (BYOL)"