The AI Scientist by Sakana AI: status and timeline

The AI Scientist

Automate scientific research so that discovery scales with computing power, starting with AI research, where every experiment runs on a computer.

ProtoStage 2 of 5

Published in Nature (Mar 2026); v1 and v2 are open source. Output is workshop-level, not top-venue.

Updated 26 Mar 2026Checked 11 Oct0 updates this week

Milestones

No announced next step

AI Scientist v1 released, about $15 per paper13 Aug 2024Complete.
AI-written paper passes ICLR 2025 workshop review12 Mar 2025Complete.
AI Scientist-v2 paper and code releasedApr 2025Complete.
Peer-reviewed paper published in Nature26 Mar 2026Complete.

Most important updates

Current obstacles

Ideas and rigour are still shallowIts own authors report naive ideas, weak methods and trouble with complex code. Output passes workshops, not main conferences.
Hallucinated detailsGenerated papers can cite sources wrongly or contain obvious mistakes, so every result still needs human checking.
Publishing norms for AI papersThe workshop paper was withdrawn before publication because venues have not agreed whether to accept fully AI-written work.

Physics limits

Language models guess, they do not knowThe system is built on models that predict likely text. They can produce fluent but false claims, and nothing inside them guarantees a result is true.
Only experiments a computer can runIt works for machine learning because the whole experiment is code. Chemistry or biology needs physical lab work, which is slow and costly to automate.
Checking is the bottleneckProducing papers is cheap; verifying them is not. An automated reviewer can miss the same flaws the writer made, so human review effort barely shrinks.

How it works

4 parts

Ideas

Generate and screen ideas

A language model proposes research ideas, then searches the literature to drop ones already published.

Experiments

Tree search over experiments

A manager agent writes and runs code, branching into variants and keeping the ones that work, like a researcher trying many runs.

Writing

Writes the paper in LaTeX

It writes the full paper, and a vision model checks the figures so they read correctly.

Review

Automated reviewer

A separate model scores each paper the way a conference reviewer would, giving feedback for revision.

Spec sheet

Spec	The AI Scientist
Compute cost per paper (v1)	About $15^R (reported)
Workshop review score (v2)	6.33 average (6, 7, 6)^R (reported)
Papers sent to ICLR workshop	3 (1 passed review)^R (reported)

^R reported by the company

Papers & demos

2 items

Apr 2025paper
The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search (opens arxiv.org)Dropped human code templates and added tree search over experiments; produced the paper that passed workshop review.
13 Aug 2024paper
The AI Scientist: towards fully automated open-ended scientific discovery (opens sakana.ai)First version: generated ML papers end to end for about $15 each, scored by an automated reviewer.

Update log

3 updates

About Sakana AI

The team behind The AI Scientist

Sakana AI

AI Scientist, Japanese enterprise AI

Tokyo AI lab founded in 2023 by former Google researchers David Ha and Llion Jones with Ren Ito. It builds AI tuned for Japanese language and business, and its AI Scientist system automates machine-learning research end to end.

Its AI Scientist paper ran in Nature in March 2026, after one of its fully AI-written papers passed review at an ICLR 2025 workshop.
Raised ¥20B ($135M) in a November 2025 Series B at a $2.65B post-money valuation; about $379M raised in total.
Paying work comes from banks (MUFG, SMBC, Daiwa Securities) and from Japan's defence ministry and its research agency.

sakana.ai (opens sakana.ai)Company profile →

Founded: 20233 yrs
Headquarters: Japan
Status: Private
Valuation: $2.6BprivateNov 2025
Raised: $135M1 round
Last round: Series B · $135M17 Nov 2025
Works in: AIAI for science
Coverage: 1 program · 4 updateslatest 26 Mar 2026checked 11 Oct
Customers: MUFG (opens mufg.jp)Daiwa Securities GroupJapan Ministry of Defense
Investors: Khosla Ventures
People: David Ha (opens sakana.ai)Co-founder & CEOLlion Jones (opens sakana.ai)Co-founder & CTORen Ito (opens sakana.ai)Co-founder & Chairman

Spec

The AI Scientist

Compute cost per paper (v1)

About $15^R (reported)

Workshop review score (v2)

6.33 average (6, 7, 6)^R (reported)

Papers sent to ICLR workshop

3 (1 passed review)^R (reported)

The AI Scientist

Milestones

Most important updates

Current obstacles

Physics limits

How it works

Generate and screen ideas

Tree search over experiments

Writes the paper in LaTeX

Automated reviewer

Spec sheet

Papers & demos

Update log

Thu 26 Mar

Wed 12 Mar 2025

Tue 13 Aug 2024

About Sakana AI

Current obstacles

Physics limits

How it works

Generate and screen ideas

Tree search over experiments

Writes the paper in LaTeX

Automated reviewer

Spec sheet

Papers & demos

Update log

Thu 26 Mar

Wed 12 Mar 2025

Tue 13 Aug 2024

About Sakana AI