Retrieved article excerpt
Open article · Retrieved 2026-09-10T16:24:29.360755+00:00
? Talleyrand GitHub Sign in An AI workspace for research & thinking Ask better questions Research means dozens of open questions, and it's easy to lose the thread. Talleyrand keeps them in one tree, always shows you the next question worth asking. The longer you work, the better it understands you. Open a case See how it works It's all free and open-source , but you need your own OpenAI key (Anthropic optional, for Claude answers). talleyrand · research AI capabilities timelines Report Brief Forecast AI capabilities through 2030 to work out which policies governments should adopt. When could AI automate AI research? How to measure R&D automation? Why does AI feel faster than it is? Will the 7-month doubling hold? Why do agents fail at long tasks? How binding are compute limits? What did past forecasts get wrong? ‹ › Automating research › Measurement How would we actually measure AI R&D automation? Not by asking people: in the only randomized trial so far, developers using AI felt 20% faster but measured 19% slower. The longest task AI can finish alone has doubled every four months . Before extrapolating, see ↳ Why do agents fail at long tasks? Suggested questions If AI were already speeding up research inside labs, what would we see from outside? Ask a follow-up question… See it in action Research in four moves Talking to LLMs is great for deep topics, but linear chat isn't the best UI. It grows, and you end up scrolling back and forth. In Talleyrand every question stays in a subtree, the AI reads all of it, and every reaction you give teaches it what you want. 01 Start from a mess, not a blank page This is how you kickstart the research. Type or dictate a stream of thoughts — what you already know, what you want to know, notes, doubts, half-baked thoughts. Talleyrand turns the mess into a short brief and your first questions. Your brain-dump I'm a policy analyst. I know export controls and the EU AI Act well. But I can't tell whether compute governance actually slows dangerous AI capabilities or just moves them elsewhere. Some people I trust say one thing, some say the opposite. I want to know which is true. Draft brief & questions Your brief and starting questions will appear here. 02 Branch a follow-up from any sentence See something worth digging into? Select it and ask. Every answer is written with the entire case in context: the brief, every thread, your highlights and reactions. Is moral progress real, or just moral change? Short answer: yes for abolition, on a standard the losers could have accepted; no for "the moral circle expands" as a general law. First, deflate the deflation. "Every generation thinks it's more moral" is true of science too, and there the inference from confident to deluded obviously fails. The pattern of self-congratulation is evidence of bias, not evidence of no fact. ↑ Select the highlighted phrase to branch 03 A question tree that tracks what you've answered Your whole case lives in one tree of questions. And it's not just for you: Talleyrand knows what you've read, what's waiting, and what's settled, and takes that into account. What do Big Five traits actually predict? Is a correlation of .3 big or small? Why do effect sizes shrink on replication? Do traits change over a lifetime? Is Myers-Briggs any better? What do Big Five traits actually predict? Effects are real and mostly replicate, but they live in the r = 0.1–0.3 band. … Two structural facts worth holding onto. First, the workhorses are Conscientiousness and Neuroticism; Openness and Agreeableness predict tastes and relationships more than achievement. Open this case → 04 The AI proposes what's next and learns from your reactions After each answer, Talleyrand proposes what to ask next, plus big-picture violet questions that challenge your framing. Accept them, or decline and say why. Every heart, highlight, and decline improves its understanding of your intent. The longer a case runs, the better it gets. Try everything — heart the answer, flag the underlined phrase, accept or decline the questions. … Determinism is not an experimental finding. If universal physical determination threatens free will, milliseconds add nothing; if it doesn't, they still add nothing. Suggested questions What experiment could distinguish a genuinely settled decision from mere neural bias or preparation before action? ? How robust is the finding that readiness potentials disappear for deliberate, consequential choices, and what are its main methodological weaknesses? ? What does roughly 60% advance prediction actually show once temporal leakage, choice-history effects, and out-of-sample replication are accounted for? ? Talleyrand's read on you A new case — still learning what you're after. Open this case → And the rest of the desk Everything the deep dive needs, in one place Insight highlights Flag a passage with the lightbulb; those insights improve every later answer and suggestion. Cross-links Answers reference other questions as live chips — click one to jump straight to that thread. Choose your model Answer with the latest GPT or Claude models at any thinking level. Easily regenerate with another. Attach documents Give the case or a single question its own source material to reason over. Love the best answers Heart the answers that land. They become a strong signal for the report and later prompts. One-click report Turn the whole case into a clean Markdown write-up you can copy or download. Share read-only Publish a case as a link. Readers can follow the whole tree, and copy the case to continue the work themselves. Dictate a question Think out loud — speak your question and it's transcribed straight into the composer. Runs while you are away Answers generate on the server, so closing the tab never loses in-flight work. ? Stop losing your train of thought in a scroll of chat. Give a messy question room to branch. Talleyrand keeps the structure so you can keep thinking. Start free with your own key We would like to use Google Analytics to count visits and see which parts of Talleyrand get used. It stores a cookie in your browser, and never receives your questions, answers, or documents. You can change your mind at any time. Details are in the privacy policy . Decline Accept