PUBLISHED AI RESEARCH AND PROJECTS

AI research and projects

The research, projects, and explanations behind the people. Read the source, inspect verified code and models, and find a collaborator.

A source check confirms that a source supports the summary. It is not an independent replication or endorsement. Our methodology ↗

Showing 49–64 of 567 proofs of work · Newest first
Page 4 of 36
researchChecked Sep 23, 2026

A Neural Probabilistic Language Model

Yoshua Bengio · Mila - Quebec AI Institute / Université de Montréal

Introduced the seminal concept of learning continuous distributed word feature vectors alongside statistical language modeling, defeating the curse of dimensionality and establishing modern neural language modeling.

Scope & limitations

Computational bottleneck in softmax normalization over large vocabularies prior to hierarchical softmax and negative sampling.

Context: Feed-forward neural network sharing distributed representation layers across context words.

View mission
researchChecked Sep 23, 2026

Adversarial Attacks on Neural Networks in Physical and Digital Environments

Dawn Song · UC Berkeley

Demonstrated that adversarial perturbations crafted via gradient-based optimization survive physical-world transformations (such as camera capture, lighting variations, and printing), exposing critical safety vulnerabilities in autonomous perception systems.

Scope & limitations

Defending against physical attacks often degrades clean-sample classification accuracy on subtle visual nuances.

Context: Inception and ResNet vision models subjected to physical adversarial stickers and perturbations.

View mission
researchChecked Sep 23, 2026

AI Feynman: A Physics-Inspired Method for Symbolic Regression

Max Tegmark · MIT / Future of Life Institute

Developed AI Feynman, a recursive neural network algorithm combining dimensional analysis, symmetry detection, and polynomial decomposition to discover exact symbolic mathematical laws from raw scientific experimental data.

Scope & limitations

Exponential complexity scaling in high-dimensional systems without clear separability or symmetries.

Context: Neural networks trained to identify translational, rotational, and scaling invariants across 100 equations from the Feynman Lectures on Physics.

View mission
implementationChecked Sep 23, 2026

Air Street Capital AI Spinout & Tech Bio Investment Playbook

Nathan Benaich · Air Street Capital

Created the venture evaluation blueprint helping top European and American academic research spinouts commercialize foundation models in biology, materials discovery, and physics simulations.

Scope & limitations

TechBio AI models require expensive wet-lab validation cycles that lag in silico prediction speeds.

Context: Deep learning models applied to protein folding, molecular generative dynamics, and automated wet lab integrations.

View mission
explanationChecked Sep 23, 2026

Algorithmic Injustice: A Relational Ethics Approach to AI Systems

Abeba Birhane · Mozilla / Trinity College Dublin

Formulated the foundational relational cognitive framework critiquing the reduction of human social complexity into static statistical classification matrices, demonstrating how predictive AI disproportionately harms marginalized populations.

Scope & limitations

Requires regulatory and legislative enforcement to compel commercial organizations to adopt rigorous external auditing.

Context: Applied across algorithmic criminal sentencing, hiring filters, and welfare fraud scoring systems.

View mission
implementationChecked Sep 23, 2026

Anthropic Responsible Scaling Policy (RSP) Framework Version 1.0 & 2.0

Daniela Amodei · Anthropic

Formulated the industry-standard risk management framework linking model capability thresholds (ASL-1 through ASL-4) directly to mandatory physical cybersecurity controls, red-teaming verifications, and operational shutdown protocols prior to model deployment.

Scope & limitations

Requires continuous empirical evaluation pipelines to reliably detect latent dangerous capabilities in reasoning models before public exposure.

Context: Governance protocol spanning internal cluster security, model weight custody, and chemical/biological risk mitigation across Claude foundation models.

View mission
implementationChecked Sep 23, 2026

ARC Autonomous Replication & Threat Evaluation Framework (METR)

Paul Christiano · Alignment Research Center (ARC)

Created the standardized red-teaming benchmark testing whether frontier AI models possess autonomous cyber-offense capabilities, resource acquisition, self-replication, or model weight exfiltration skills.

Scope & limitations

Agent capabilities evolve rapidly with multi-agent orchestration, requiring continuous suite updates.

Context: Sandboxed agentic execution environments testing GPT-4, Claude 3, and open weights models.

View mission
implementationChecked Sep 23, 2026

Build a Large Language Model (From Scratch): Complete PyTorch Implementation

Sebastian Raschka · Lightning AI

Created the canonical open-source educational repository teaching engineers how to code, train, and fine-tune a complete GPT-style transformer from scratch in pure PyTorch without external high-level libraries.

Scope & limitations

Designed primarily for educational transparency; industrial scale multi-node training requires distributed FSDP/Megatron wrappers.

Context: Modular Python/PyTorch codebase implementing multi-head self-attention, BPE tokenization, causal masking, and instruction alignment.

View mission
researchChecked Sep 23, 2026

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Hyung Won Chung · OpenAI

Co-authored the foundational paper discovering that generating step-by-step reasoning tokens before the final answer dramatically unlocks multi-step arithmetic, commonsense, and symbolic reasoning in foundation models.

Scope & limitations

Chain-of-thought benefits appear as an emergent property only in models with sufficient scale (~100B+ parameters or instruction-tuned compact models).

Context: PaLM 540B, GPT-3 175B tested on GSM8K math benchmark and SVAMP.

View mission
implementationChecked Sep 23, 2026

ChatGPT Production Launch & Reinforcement Learning from Human Feedback (RLHF) Scaling

Mira Murati · Independent AI Research Lab

Directed the technical conversion of GPT-3.5 into ChatGPT via conversational RLHF, dialog-state conditioning, and distributed inference optimization that scaled to 100 million active users in two months with high system availability.

Scope & limitations

Early versions suffered from sycophancy and hallucinations under adversarial jailbreak prompts, necessitating iterative red-teaming.

Context: PPO reinforcement learning on InstructGPT checkpoints deployed across large-scale NVIDIA A100 GPU clusters.

View mission
implementationChecked Sep 23, 2026

Confidential Computing & Privacy-Preserving LLM Inference Architectures

Dawn Song · UC Berkeley

Engineered secure enclave and hardware-assisted cryptographic protocols ensuring user prompt embeddings and model weights remain provably encrypted during remote cloud GPU inference.

Scope & limitations

Enclave memory constraints introduce bandwidth overhead during large context-window KV cache allocation.

Context: NVIDIA H100 confidential computing enclaves and AMD SEV-SNP virtual machine attestation.

View mission
researchChecked Sep 23, 2026

Constitutional AI: Harmlessness from AI Feedback (RLAIF)

Dario Amodei · Anthropic

Introduced the Constitutional AI paradigm enabling models to critique, revise, and align their own behavioral outputs according to written principles without human-labeled toxicity feedback. Established the foundational technique powering Claude's red-teaming resilience.

Scope & limitations

Constitution design must be continuously audited for latent political biases, conflicting clauses, and nuanced real-world legal edge cases.

Context: Frontier autoregressive transformer checkpoints trained via self-critique, Chain of Thought, and Reinforcement Learning from AI Feedback (RLAIF).

View mission
researchChecked Sep 23, 2026

Deep Reinforcement Learning from Human Preferences (RLHF)

Paul Christiano · Alignment Research Center (ARC)

Authored the foundational paper establishing that training deep neural policies via pairwise human preference reward models allows agents to master complex tasks (Atari games and robotic backflips) without hard-coded programmatic reward functions.

Scope & limitations

Reward models can be gamified by agents via reward hacking and sycophantic responses.

Context: PPO actor-critic architectures and learned Bradley-Terry reward models.

View mission
researchChecked Sep 23, 2026

Dota 2 with Large Scale Deep Reinforcement Learning (OpenAI Five)

Greg Brockman · OpenAI

Led the engineering and systems team that built OpenAI Five, scaling Proximal Policy Optimization (PPO) across 128,000 CPU cores and 256 GPUs playing 180 years of game simulation daily, defeating reigning world champion esports team OG.

Scope & limitations

Required heavy reward engineering and hero draft restrictions to converge on winning policies in complex game spaces.

Context: 4,096-unit single-layer LSTM processing continuous state-action spaces with 20,000-step time horizons.

View mission
explanationChecked Sep 23, 2026

Enterprise Alignment & Secure Frontier Deployment Standards

Daniela Amodei · Anthropic

Architected Anthropic's enterprise data confidentiality guarantees, zero-data-retention training policies, and SOC2 Type II compliance frameworks adopted by major healthcare, financial, and legal institutions worldwide.

Scope & limitations

Enterprise zero-retention mandates preclude autonomous model feedback loop improvements from raw user traffic.

Context: Claude Enterprise platform on AWS Bedrock and Google Cloud Vertex AI.

View mission
researchChecked Sep 23, 2026

Eureka: Human-Level Reward Design via Coding Large Language Models

Anima Anandkumar · California Institute of Technology (Caltech)

Created Eureka, an autonomous algorithm leveraging frontier coding LLMs to synthesize, evaluate, and iteratively refine complex reward algorithms for reinforcement learning, outperforming human expert-designed rewards across 29 dexterous robot manipulation tasks.

Scope & limitations

Requires fast GPU simulation environments to provide high-throughput performance feedback to the outer-loop LLM.

Context: GPT-4 reward coding loops interacting with NVIDIA Isaac Gym GPU physics simulations.

View mission
Showing 49–64 of 567 proofs of work