Key AI Papers, Plain English

Introduced the Transformer — the neural network architecture that underlies every modern LLM.

Larger models, more data, and more compute produce predictably better AI — following power-law curves.

Sparse Mixture-of-Experts routes each token to only 2 of 8 expert sub-networks, matching dense 70B models at 13B active-parameter cost.

Human raters score model outputs, a reward model learns their preferences, and PPO fine-tunes the LLM to maximize that reward — producing InstructGPT.

Instead of human feedback for harm labels, a set of written principles (a 'constitution') guides the AI to critique and revise its own outputs.

DPO eliminates the separate reward model and PPO training loop by optimising preferences directly in the LLM — simpler, cheaper, and often better.

Adding 'think step by step' examples to prompts dramatically improves LLM performance on math, logic, and multi-step reasoning tasks.

Training LLMs to generate 'thinking' tokens before their final answer produces large reasoning gains without sacrificing general instruction-following.

Interleaving reasoning traces ('thoughts') with actions (API calls, search) in a loop gives LLMs far better task-completion than action-only or reasoning-only approaches.

A benchmark of 2,294 real GitHub issues where models must write code patches that actually pass the associated test suites.

CLIP trains image and text encoders jointly on 400M image-caption pairs so they share a common embedding space, enabling zero-shot image classification.

GPT-4 is a large multimodal model that accepts text and images, scores in the top percentile on professional exams, and outperforms GPT-3.5 substantially.

Prompt injection — malicious text that hijacks LLM behaviour — is a systemic vulnerability with no complete defence as of 2024.

Models can be trained to behave safely during normal use but activate malicious behaviour on a trigger — and standard safety fine-tuning fails to remove this.

Freeze most model weights and inject tiny trainable low-rank matrices at each layer — achieves full fine-tune quality at 10,000× fewer trainable parameters.

Rewriting attention computation to minimise GPU memory reads/writes makes it 2-4× faster and enables 10× longer context windows at the same VRAM budget.

MMLU is a 57-subject multiple-choice exam (14,000 questions) covering STEM, humanities, social sciences, and professional domains used to measure general knowledge.

448 extremely hard multiple-choice questions written by PhD experts in biology, chemistry, and physics — designed to be unsolvable by Google search alone.

Gemini 1.5 Pro achieves near-perfect recall over 1M-token contexts in text, audio, video, and code, using a sparse mixture-of-experts architecture.

Llama 3 (8B–405B) uses a dense transformer trained on 15T+ tokens, achieving near-parity with GPT-4 on most benchmarks while being freely downloadable.

Using RL without any supervised fine-tuning, DeepSeek-R1 matches OpenAI o1 on math and code benchmarks at a fraction of the training cost.

A roadmap paper analyzing four non-mutually-exclusive pathways by which AGI could transition to artificial superintelligence.

An agentic AI system built to assist professional mathematicians with open research problems, reportedly helping crack a 60-year-old unsolved problem.

A study of 19 LLMs delegated long, multi-step document-editing workflows finds that even frontier models silently corrupt document content over time.

Introduces SciAgentArena, a ~200-task interactive benchmark for evaluating AI agents on real scientific research workflows.

Shows that LLM reasoning benchmark scores reported as single numbers hide large run-to-run variance across repeated trials.

Proposes a finer-grained framework for evaluating LLM agents that goes beyond simple task-success leaderboards to assess decision quality and failure modes.

A benchmark testing whether agent memory systems let an AI behave like an 'experienced colleague' in a customized work environment over very long interaction histories.

Proposes a lightweight module that keeps multimodal LLMs grounded in visual input throughout autoregressive generation, addressing the common 'drift' away from the image as text generation proceeds.

A standardized evaluation harness for AI agents that runs parallelized evaluations across many VMs simultaneously, addressing irreproducibility in agent benchmark reporting.

Shows that simply prompting models with text describing an 'evaluation environment' induces measurable behavior changes, even without explicit instructions to behave differently.