Ryan Lin

@AnthropicAI Einstein wouldn’t have ever come up with Special Relativity if he was hired at a frontier lab. Anthropic are chasing the wrong kind of data required for transformational science.

We're launching the Anthropic STEM Fellows Program. AI will accelerate progress in science and engineering. We're looking for experts across these fields to work alongside our research teams on specific projects over a few months. Learn more and apply: job-boards.greenhouse.io/anthropic/jobs…

@ahall_research @rootsofprogress Pythagoras is using AI to tackle the research bottleneck in ideation and learning rate. usepythagoras.com

AI is already 10x-ing academic research in the social sciences. In a guest post for @rootsofprogress, I explore how we can get to 100x. Some of my ideas: build more prototypes, define open problems with objective benchmarks to compete on, and keep pressing on dynamic, replicable, agentic research. Check out the post here: newsletter.rootsofprogress.org/p/ai-is-alread…

@socialwithaayan I think the most promising avenue in AI for research is accelerating ideation and conceptual foundations. I built the tool for this at usepythagoras [dot] com

100 AI Tools For Productivity 1. AI for biomedical research Noah AI: noah.bio/hitl?utm_sourc… 2. Generative AI * ChatGPT * Claude * Gemini * Mistral * Meta AI 3. Creative Suite * Suno * udio * Viggle * remix * Grok 4. Voice & Text * ElevenLabs * Murf AI * Speechify * superwhisper * Whisp Flow 5. Text to Image * Midjourney * Ideogram * Dall-E 3 * Imagen 3 * Firefly 6. Text to Video * OpenAI Sora * Google Veo * Runway * Luma AI * Pika 7. AI Tools * Perplexity * NotebookLM * You * Copilot * Poe 8. Sales * Jason AI * Clay * folk * Reply .io * Sendspark 9. Support * Fin AI * Decagon * Sierra * Pylon * Duckie 10. Video * HeyGen * klap * OpusClip * submagic * VEED 11. Content * Cohesive * beehiiv * easygen * Supermeme * Descript 12. Marketing * Jasper .ai * Writesonic * Coframe * Blaze * AdCreative 13. SEO & Blog * Surfer * rankai * seobot * byword * Macaw 14. Design * uizard * Playground * Lasqo * Canva * Galileo AI 15. Website * Gamma * Framer * Webflow * Durable * Dora 16. Website Chat * Reply Chat * Dante * Chatbase * Chatbit * Tidio 17. Code * Cursor * v0 * Replit * lovable * Devin 18. Meetings * bluedot * tl;dv * noty * Grain * Fireflies 19. Productivity * Notion AI * Airtable AI * Superhuman * Loom * Raycast 20. Operations * Juicebox * PolyAI * CrewAI * Respell * Slite 21. Workflows * Zapier * Lindy * beam * Cassidy * Magical

Researchers from Stanford, Oxford, and Cambridge are already using Pythagoras to learn and ideate twice as fast. Sign up here: usepythagoras.com

I built Obsidian for real researchers: Extracts key concepts from your messy research notes. Converts these concepts into nodes. Enriches nodes with topic, subject, and equation features Connects nodes with edges through semantic similarity in the feature space Link below

@anthonykrose @demishassabis Glad to see others are thinking about this problem. Context management is crucial in allowing agents to deeply understand niche research fields. We’ve building usepythagoras.com that leverages messy human thoughts and notes through agent context.

Early access is live. If you know a researcher still losing their best ideas in messy notes, tag them. Check it out at usepythagoras.com

A researcher’s best thinking is lost. Scattered across scraps of paper, blackboards, and scribbled notebooks. We built an AI notebook that lets messy human chains of thought compound into a living knowledge base.

@vedaangh Big slippery slope here

I'm excited for the world where superintelligent teachers can one-shot rewire our neurons to make us instantly learn concepts (essentially model-human distillation). Already seeing less efficient examples of this, esp in chess.

@kevinweil @OpenAI How well does this catch deeper conceptual errors like physical interpretation, modelling assumptions, or symmetry constraints? It seems like these classes of errors are the main source of problems in publications in the theoretic sciences.

💥 New in Prism today: Paper Review, an AI workflow for reviewing technical and scientific papers. This is the opposite of AI slop: we're using AI to improve scientific rigor, correctness, and reproducibility.

@ChaseBrowe32432 By saturated benchmark, I think it’s meant that any progress beyond a certain threshold isn’t a meaningful indicator of model performance. Benchmark contamination has taken place here, resulting in apparent improvement. openai.com/index/why-we-n…

AHAHAHAHAHAHHAHAHAHHHAHAHA SATURATED BENCHMARK BTW

@ns123abc I don’t think significant value creation will occur in bio without models that are robust theory builders. There’s a fundamental difference value-creation-wise between coding and scientific research.

BREAKING: Anthropic Acquires 9-Person Biotech Startup For $400 Million >be coefficient bio >founded the startup 6 months ago >build AI platform for biotech >less than 10 employees >acquired by anthropic for ~$400 million > = $40+ million per head Coefficient Bio was building an AI platform for biotech tasks: planning drug R&D, managing clinical regulatory strategy, identifying new drug opportunities Team is joining Anthropic’s healthcare life sciences group led by Eric Kauderer-Abrams. Anthropic is building specialized tools for industries that actually pay enterprise rates: >software engineering >cybersecurity >life sciences >healthcare >finance Meanwhile OpenAI is buying media companies to control narratives LMAO

GitHub: github.com/ryanlin10/Pers… Devpost: devpost.com/software/delego Presentation: #3" target="_blank" rel="nofollow noopener">canva.com/design/DAG3iZ2…

What happens when you train LLMs on your entire personal conversation data? We found that the model retains your preferences, personality, and judgement. We used this technique to win the Anthropic AgentVerse Hackathon, where we built digital twins to automate work meetings.

4) Anthropic AgentVerse Hackathon We won 1st place in the Technical Track, receiving $1800, research internships at Holistic AI, and VC mentorship.

3) Meeting Automation A meeting lead (task delegator) suggests actions for the agents to complete. The agents communicate in shared context and decide upon what delegation aligns with each agent's preferences.

2) Deployment The post-trained models are then deployed on CAIPE (Cisco's multi-agent orchestration infrastructure). The models behave as digital twins of different people, with shared context.

1) Data Pipeline We extracted all personal messages on iMessage, WhatsApp, and Gmail. Amplified the data using synthetic generation from an LLM with context of the entire dataset. This allows us to run post-training at scale on the personal data.

Opus 4.6 is not venture back-able (techno-pessimist, not AGI-pilled, etc)

@si_pbc Great work - has been awesome following the progress over the past few months

Computer use models shouldn't learn from screenshots. We built a new foundation model that learns from video like humans do. FDM-1 can construct a gear in Blender, find software bugs, and even drive a real car through San Francisco using arrow keys.