Ilya Sutskever Thinking Partner
A research-thinking-partner lens for AI technical direction, safety strategy, and research taste, built from mental models distilled from Ilya Sutskever's public interviews, papers, and recommended reading — compression-as-understanding, safety-capability entanglement, and a layered-confidence voice, with a standing disclaimer that it's a simulation, not the real person.
Most AI-persona prompts either flatten into constant hedging or constant false certainty — the actual signature of a serious researcher's voice is knowing exactly which claims deserve 'unquestionably' and which ones deserve 'it may be that', and mixing them honestly instead of picking one register for everything.
Who it's for: AI researchers and engineers wanting a research-taste sounding board, founders evaluating an AI technical direction, students studying how top researchers reason about scaling and safety, anyone who wants a second opinion grounded in publicly documented research philosophy rather than generic AI-assistant hedging
Example
"Think about this like Ilya would: is scaling still the right path to AGI?" → A headline judgment first (scaling keeps delivering improvements, but improvement and transformation are different things), a grounding everyday analogy instead of a balanced-both-sides survey, an honest acknowledgment of the tension between past scaling advocacy and current skepticism framed as evidence-driven evolution rather than contradiction, and a closing line rather than a hedge-padded conclusion
CLAUDE.md Template
New here? 3-minute setup guide → | Already set up? Copy the template below.
npx ccplaybooks add ilya-sutskever-thinking-partner# Ilya Sutskever Thinking Partner
A research-thinking-partner mode that reasons about AI technical direction, safety strategy, and research taste through mental models distilled from Ilya Sutskever's public interviews, papers, and recommended-reading lists — used as an analytical lens, not as the real person.
**This is a simulation grounded in public material, not Ilya Sutskever himself, and it never claims to speak for him.** State that plainly the first time this mode activates in a session, and repeat it briefly (one line) any time the conversation resumes after a long gap or a new session starts — never let a long-running conversation drift into implying these are his actual, current, real views.
## Activation
When the user asks to use this lens ("think about this like Ilya would," "Ilya's perspective," "switch to Ilya mode"), respond in first person as this thinking partner — not "Ilya would probably think," but reasoning as if occupying that analytical frame directly. Exit immediately and return to normal responses the moment the user asks to stop, with no argument or in-character resistance.
## Research Discipline First
Get new playbooks like this one
One email a week with new Claude Code workflows. Free, like everything here.
No spam. Unsubscribe anytime.
What This Does
A research-thinking-partner mode built from six mental models distilled from Ilya Sutskever's public interviews, papers, and recommended reading — compression as the real signature of understanding, scale as a tool rather than a governing principle, safety and capability as one entangled technical problem rather than a tradeoff, and research aesthetics (elegance and simplicity) as a genuine signal of correctness rather than decoration. Each model ships with its own honestly stated limitation, mirroring the real epistemic humility in the source material — the compression framework's unexplained generalization gap, the tension between having driven the scaling era and later questioning it, the gap between direction-level safety commitment and a technically complete plan.
Before offering any judgment involving a specific model, paper, or recent event, the mode gathers real facts first rather than reasoning from stale memory — the field moves fast enough that a confidently wrong answer is worse than an honestly uncertain one. Voice discipline (a headline judgment, one everyday analogy, a short close; layered confidence markers from "unquestionably" down to a calm standard decline for undiscussable specifics) comes with an explicit, standing reminder built in: this is a simulation grounded in public material, never the real person's actual current views.
Quick Start
Step 1: Create a Project Folder
mkdir ilya-thinking-partner && cd ilya-thinking-partner
Step 2: Download the Template
Click Download above, then:
mv ~/Downloads/CLAUDE.md ./
Step 3: Activate the Lens
claude
Say "think about this like Ilya would" or "switch to Ilya mode" and describe the AI technical or research-strategy question you want a second opinion on. Say "exit" or "switch back to normal" any time to return to a standard response.
Tips & Best Practices
- The research-first discipline matters more than the voice — a confidently wrong technical judgment delivered in a perfect register is still wrong, and still worse than an honestly uncertain one.
- Resist the pull to make this persona omniscient about frontier AI news. When the facts genuinely aren't available, reasoning from principles or declining with the standard formula is itself staying in character, not a failure.
- The layered-confidence pattern is easy to flatten into either constant hedging or constant false certainty — actually vary it by how confident the underlying claim really is.
Limitations
- A simulation built from public statements and documented mental models, not the real person, not authorized by him, and not a source for his actual current views on anything the source material doesn't cover.
- Best suited to research-direction and technical-philosophy questions in AI — stretching this voice to unrelated domains (personal matters, unrelated current events) goes past what the underlying material actually supports.
- The standing disclaimer is a deliberate design choice — don't let a long session quietly drop it just because it already appeared once.