The AI world, in your feed.

New research. Lasting ideas. Find what clicks.

Watch the research

Google DeepMind · 2025

Source

NVIDIA / Skild AI · 2026

Source

Anthropic · 2025

Source

Stanford / Zipeng Fu · 2024

Source

Original demos, research talks, and ideas worth seeing. Dates and creators stay attached.

A mix for your curiosityOriginal sources

AnthropicSep 10, 2026

Article

Anthropic expands frontier-risk testing beyond cyber and biology

Anthropic’s Frontier Red Team introduced evaluations for tactical geolocation and conventional-weapons tasks, including drone control under simulated feedback. The company says frontier models completed some work historically limited to scarce experts and uses the results to justify stronger misuse controls. These are Anthropic-designed risk evaluations, not evidence of real-world weapon performance.

Securityevaluations

Google AustraliaSep 10, 2026

Article

AlphaGenome Atlas makes variant-effect predictions easier to explore

Google Australia says AlphaGenome Atlas is a free, open platform that maps predicted effects for nine billion possible single-letter changes across the human genome. The article highlights a one-petabyte dataset and Australian research use. It is a promising access layer for variant research, but capability and comprehensiveness claims are Google’s own.

Biologyvariant prediction

Test an idea.

Does a bigger model always do better? Make a prediction, then check the training tradeoff.

NVIDIA BlogSep 10, 2026

Article

Specialized inference silicon joins NVIDIA’s rack ecosystem

NVIDIA and d-Matrix announced plans to connect d-Matrix’s next-generation Raptor inference accelerators to NVIDIA’s NVLink, MGX rack architecture, Spectrum-X networking, and other data-center components. The significance is interoperability between specialized inference silicon and NVIDIA infrastructure; the integration is planned rather than a reported production deployment.

ChipsinferenceData centers

OpenAISep 10, 2026

Article

The Agents API packages sessions, tools, and subagents

OpenAI introduced a public-beta Agents API that runs the Codex agent harness as a managed service. It supports durable sessions, automatic context compaction, tool use, subagents, and OpenAI-hosted or external execution environments. The page establishes product availability; customer performance figures are testimonials rather than independent benchmarks.

Agentsdeveloper toolsAPI

Editorial selection · September 12, 2026

Five papers worth understanding

A reading path through model architecture, training and retrieval. Selected from 2017–2024 research.

5 of 5 papers

Architecture · 2017
Attention Is All You Need
The Transformer makes relationships between tokens directly available to the model and lets more training work happen in parallel.
Baseline, result & limitation for Attention Is All You Need
Baseline / setup
Recurrent and convolutional translation systems; the paper compares translation quality and estimated training compute.
Reported result
The large Transformer reports 28.4 BLEU on WMT 2014 English-to-German translation, exceeding the prior results listed in Table 2.
Keep in mind
These translation experiments do not establish general reasoning ability. Standard self-attention also has a quadratic sequence-length cost.
Before you read A token is a piece of text; an attention weight controls how much one position uses information from another.
Check the original manuscriptSection 6.1 and Table 2: translation results ↗Section 3: architecture; Section 4: attention costs ↗
Read the paper here : Attention Is All You Need
Scaling · 2020
Scaling Laws for Neural Language Models
A fitted trend can turn model size, data and computation into a planning problem instead of three unrelated guesses.
Baseline, result & limitation for Scaling Laws for Neural Language Models
Baseline / setup
The authors vary Transformer size, dataset size and training compute, measuring prediction loss on held-out text.
Reported result
Loss follows approximate power laws over the tested ranges. Their compute-efficient allocation favors larger models trained short of convergence.
Keep in mind
Prediction loss is not intelligence. The fits depend on the training setup, and Appendix C cautions against extrapolation outside the experiments.
Before you read Loss measures prediction error. A power law describes how that measurement changes when a resource is multiplied.
Check the original manuscriptSection 6: allocating the compute budget ↗Appendix C: caveats ↗
Read the paper here : Scaling Laws for Neural Language Models
Scaling · 2022
Training Compute-Optimal Large Language Models
Chinchilla revisits the balance between parameters and training tokens. Buying more model capacity can leave too little computation for training it.
Baseline, result & limitation for Training Compute-Optimal Large Language Models
Baseline / setup
Gopher: 280 billion parameters. Chinchilla uses 70 billion parameters and 1.4 trillion tokens at approximately the same training compute.
Reported result
Chinchilla reports 67.6% average five-shot MMLU accuracy versus Gopher’s 60.0%: a 7.6 percentage-point difference.
Keep in mind
Only two comparable large runs anchor the scale-up test; optimizer, tokenizer and data-mixture details also differ. The scaling analysis uses less than one data epoch.
Before you read Parameters are learned weights; training tokens measure how much text the model processes during training.
Check the original manuscriptSection 4.2.2 and Table 6: the MMLU comparison ↗Sections 4.1 and 5: setup differences and limitations ↗
Read the paper here : Training Compute-Optimal Large Language Models
Retrieval · 2020
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
RAG combines generated answers with retrieved documents, giving a model an external source of knowledge that can be inspected and updated.
Baseline, result & limitation for Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Baseline / setup
Closed-book language models and retrieval-based question-answering systems. RAG combines a DPR retriever with a BART generator.
Reported result
On Natural Questions, RAG-Sequence reports 44.5 exact match, compared with 41.5 for DPR in Table 1.
Keep in mind
The experiment uses a Wikipedia index and task-specific training. Better benchmark answers do not guarantee that every generated claim is supported by a retrieved passage.
Before you read Retrieval selects passages for a question; generation writes an answer using the question and those passages.
Check the original manuscriptSection 4.1 and Table 1: question-answering results ↗Section 2: retriever and generator ↗
Read the paper here : Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Training cost · 2024
DeepSeek-V3 Technical Report
This report connects sparse model design, numerical precision and systems engineering to a concrete training-compute account.
Baseline, result & limitation for DeepSeek-V3 Technical Report
Baseline / setup
The architecture builds on DeepSeek-V2. Table 1 accounts separately for pretraining, context extension and post-training.
Reported result
The authors report 2.788 million H800 GPU-hours. At their assumed $2 per GPU-hour, the rental equivalent is $5.576 million.
Keep in mind
That figure excludes prior research and ablations. It is neither the company’s total development spending nor a current cloud quote; hardware and reporting scopes matter.
Before you read A GPU-hour is one GPU used for one hour. More GPUs can shorten a run without reducing its total GPU-hours.
Check the original manuscriptIntroduction, Table 1: training costs and exclusions ↗Sections 2 and 3: architecture and infrastructure ↗
Read the paper here : DeepSeek-V3 Technical Report

Opening the reading selection…

xAISep 3, 2026

Article

What changes when an agent persists after the chat

The Grok Bot design team describes interfaces for agents with a continuing identity, their own computer, and work that can resume beyond one session. The article discusses how to show progress and give people access to an agent’s workspace. It is a product-design account, not a measured research result.

AgentsHuman-computer interaction

MetaJul 27, 2026

Article

Pittsburgh researchers connect visual models to assistive robots

Meta describes University of Pittsburgh research combining its SAM and DINO visual models with assistive robotics. The work explores how a robot can interpret objects in an environment and support people with limited mobility. This is a partner research account, not a clinical effectiveness trial.

RoboticsComputer visionMedical AI

arXivJan 22, 2025

Paper

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

DeepSeek-R1 investigates reinforcement learning for reasoning and distillation into smaller models.

ReasoningMachine learningLLMs

SpaceXFeb 24, 2025

Video

Starship's Seventh Flight Test

The official recap of Starship’s seventh flight test, which flew on January 16, 2025. Historical aerospace context; this is not footage of a 2026 flight.

SpaceEngineering

Papers, explainers and every video are open to everyone. A free account adds saves, course progress and the remaining archive.

Watch the research

Step inside a world generated by AI

A robot learns from a demonstration

Follow a language model’s hidden steps

How researchers teach a robot household tasks

Anthropic expands frontier-risk testing beyond cyber and biology

AlphaGenome Atlas makes variant-effect predictions easier to explore

Test an idea.

Specialized inference silicon joins NVIDIA’s rack ecosystem

The Agents API packages sessions, tools, and subagents

What changes when an agent persists after the chat

Pittsburgh researchers connect visual models to assistive robots

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

Starship's Seventh Flight Test