← all topics

🔬 Research

Frontier research talks — new architectures, training techniques, theoretical insights, paper deep-dives.

13 videos · continual-learningreinforcement-learningagiexpertiseneocognitionworld-models

The workflow

flowchart LR
    A[Open problem] --> B[Hypothesis<br/>+ experiment design]
    B --> C[Run + ablations]
    C --> D[Compare to<br/>strong baselines]
    D --> E{Holds up?}
    E -->|No| B
    E -->|Yes| F[Write-up +<br/>code release]

The cutting edge — usually 6-18 months ahead of production.

Key takeaways

Meta's Code World Model predicts program execution traces as an autoregressive sequence so agents can imagine outcomes before running code.
Scaling RL is now a talent and tooling problem; opening up RL environments and infra is how Prime Intellect plans to widen the researcher pool.
AGI's hardest problems (memory, alignment, deception, idioms, hive-mind) map nicely onto sci-fi memes — and graph-based grounding is one tool worth taking seriously.
AGI measurement needs interactive game-based benchmarks with hidden test sets so model intelligence can't be confused with memorized training data or developer-injected priors.

Videos (13)

Intelligence + Continual Learning = Expertise — Yu Su, NeoCognition

Agents need continual learning about their specific operating environments to turn general reasoning ability into efficient expertise.

11.7K views · Aug 12, 2026

Code World Model: Building World Models for Computation – Jacob Kahn, FAIR Meta

Meta's Code World Model predicts program execution traces as an autoregressive sequence so agents can imagine outcomes before running code.

11.6K views · Dec 17, 2025

RL Environments at Scale – Will Brown, Prime Intellect

Scaling RL is now a talent and tooling problem; opening up RL environments and infra is how Prime Intellect plans to widen the researcher pool.

9.6K views · Dec 09, 2025

Reinforcement Learning without Verifiable Rewards — Will Brown, Prime Intellect

Extending RL to real work requires better feedback and reward design beyond deterministic correctness checks.

4.3K views · Jul 31, 2026

How Autoresearch is changing ML research — Zhengyao Jiang, Weco

Autoresearch amplifies disciplined execution of existing ideas, making evaluation and search-space design especially valuable.

3.2K views · Jul 16, 2026

First Steps Toward Automated AI Research — Richard Socher, CEO Recursive AI

Automated scientific discovery should combine exploration with empirical selection rather than rely on idea generation alone.

2.7K views · Jul 30, 2026

Einstein Arena: Harnessing Collective Agent Intelligence for Open Science — James Zou, Together AI

Well-designed environments with verifiable rewards and shared discoveries can amplify collective scientific problem solving.

1.9K views · Aug 25, 2026

From RL to IRL — Gaurav Mishra, Amazon AGI Lab

Computer-use agents need training environments that teach recovery and judgment under real-world failure conditions.

1.8K views · Aug 14, 2026

Task Fidelity Scaling Laws — Kobie Crawdord, Snorkel

Agent training tasks should be difficult for meaningful reasons, with reliable environments that produce useful learning signals.

1.6K views · Jun 02, 2026

Adaption Labs: Gradient-Free Continual Learning — Sara Hooker, Adaption

Automating the full data-and-model adaptation loop can make specialized AI development more accessible and compute-efficient.

1.4K views · Aug 12, 2026

Scaling Compute on Context — Jack Morris, Engram

Scaling computation over private context is a proposed route from broad pretrained knowledge to personalized depth.

930 views · Aug 12, 2026

Top Ten Challenges to Reach AGI — Stephen Chin, Andreas Kollegger

AGI's hardest problems (memory, alignment, deception, idioms, hive-mind) map nicely onto sci-fi memes — and graph-based grounding is one tool worth taking seriously.

842 views · Jul 22, 2025

Measuring AGI: Interactive Reasoning Benchmarks for ARC-AGI-3 — Greg Kamradt, ARC Prize Foundation

AGI measurement needs interactive game-based benchmarks with hidden test sets so model intelligence can't be confused with memorized training data or developer-injected priors.

491 views · Jul 16, 2025