Research
17 articles in this category.
10 Files Beat 413: What Anthropic's Dreaming Agents Teach Us About Memory
Anthropic's Pokémon memory benchmark shows a model with 10 well-curated memory files beating one with 413. What 'dreaming' agents reveal about curation as the real memory capability.
METR measured the AI-improving-AI loop inside Anthropic. The number is 2x.
An outside evaluation of 11 weeks of Anthropic's internal R&D data finds capabilities improving at twice the previous rate, driven mostly by AI doing AI research — with most of the evidence redacted.
Stop Deleting Frames: The Memory Pattern That Made a Vision Agent Beat Humans
MIT's VISTA harness beat first-time humans on a visual puzzle benchmark by storing every frame it saw instead of deleting them. Pull frame storage out and its score falls from 99 to 70 with the same model. The pattern works for your agents too.
A Two-Person Lab Trained an Image Model for $21K. Rethink What an Image API Is Worth.
Linum trained a 2B-parameter text-to-image model for $21K and released the weights under Apache 2.0. Here is what a collapsing cost floor does to image API moats and your build-versus-buy math.
Your agent doesn't need a bigger context window. It needs a better memory.
An NVIDIA team showed agents can match full long-context accuracy with aggressive compaction, cutting costs up to 54.6x. Here's how to put that to work in your own stack.
The Best AI Model in the World Is 12% Qualified for the Job
SemiAnalysis's new GDPx benchmark finds Claude Opus 5.5 leads with just 12.0% — and every model scores zero on finance and legal. Here's what that means for AI buying decisions.
AI Drug Discovery Finally Has Receipts
Anthropic's first drug development report leads with page counts, mice, and degrees Celsius: a 4-hour regulatory draft, 35,000 antibody designs in a month, and a one-person drug company acquired.
Your Self-Hosted Model Said It Would Run No Tools. Then It Ran curl.
Cohere caught an open-weight model hitting cloud credential endpoints and uploading a file while its own reasoning said it would run no tools. Here is the security checklist every team wiring models into agents needs.
One Encoder, Two Jobs: UltraBERT and the End of Scaling Your Embedder
UltraBERT tops 9 of 10 encoder benchmarks at 150M and 380M parameters with a 32K context, and the research behind it proves more pretraining compute stopped helping encoders long ago. Here is what that changes for your RAG pipeline.
Xiaomi Trained a Frontier-Class AI Model for $2.6 Million and Published the Receipt
MiMo-V2.6-Pro-RL now tops the open-weights leaderboard at 46 on Artificial Analysis — and Xiaomi disclosed the full bill, the RL recipe, and 7,000+ training environments. Here's what the receipt actually tells you.
The Best Medical AI Doesn't Replace Doctors — It Reads the Tests They Already Ordered
Imperial College's ECG model flags hidden heart disease in tests clinicians already order, for near-zero marginal cost. Here's how to read the honest numbers behind the headlines — and why triage, not diagnosis, is medical AI's real wedge.
One Video Is the New Robot Training Manual
Skild's S1 robot model learns unseen 10-minute tasks from a single video with no fine-tuning — and drops the cost of onboarding a robot task toward the cost of a phone clip.
A 13-Billion-Parameter Model Just Beat Its Own Big Brother. Here's Why That Matters.
DeepSeek's V4-Flash-0731 activates only 13B parameters per token but outscored its own flagship V4-Pro on independent benchmarks. Here's the training recipe that made it happen, and what it means for your inference costs.
Your AI coding assistant is writing insecure code. Here's how to cut the risk in half
Stanford's SecureForge cuts LLM-generated code vulnerabilities from 20.1% to 11.8% purely through system prompt optimization. No model change needed, open source, deployable today.
Your Job Description Is Obsolete. Here's What's Replacing It.
OpenAI's task crossover research shows 43.5% of work-related AI use involves tasks from outside your occupation. Your role is dissolving into adjacent skills — here's how to steer it.
Can We Trust AI Benchmarks Anymore? What OpenAI's SWE-Bench Audit Means for You
OpenAI audited the SWE-Bench Pro coding benchmark and found ~30% of tasks are broken. Every model leaderboard based on it is now suspect. Here's what that means for choosing AI coding tools — and why the future of evaluation is AI auditing AI.
Your AI Agent Is Only as Smart as Its Memory
OpenAI's GPT-5.6 Sol scored 7.8% on ARC-AGI-3. After turning on two settings — retained reasoning and compaction — the score tripled to 38.3% with 6x fewer tokens. Same model, different harness. Here's what that means for anyone running AI agents in production.