Autoscience

Executive Summary

Autoscience is a San Mateo applied research lab automating machine-learning research end-to-end, under the tagline "Automate AI Research."^[1][2] Its public identity rests on two agents. Carl generates novel research hypotheses, runs experiments, and writes academic papers — Autoscience billed it as "the first AI system to produce academically peer-reviewed research" after Carl-generated papers were accepted through double-blind review at ICLR 2025 workshops, and the company has since reported Carl producing its first full-length peer-reviewed paper.^[3][4][5] Mira is the commercial turn: an ML research agent that reads 1,200+ papers a week, matches techniques to a customer's repositories, and implements improvements directly into production models — "the full pipeline from hypothesis to deployed improvement."^[6][2]

The company announced a $14M seed led by General Catalyst, with Toyota Ventures, Perplexity Fund, S32, and MaC Ventures, on March 19, 2026 — one year after the Carl milestone.^[2] That milestone is also the company's controversy: Autoscience (like rival Intology) submitted AI-generated papers to ICLR workshops without informing organizers, drew public accusations from academics of co-opting peer review for publicity, and withdrew the accepted papers.^[7] Between the two events, an Autoscience system earned a silver medal in the Kaggle Santa 2025 competition against 3,300 teams — claimed as the first time a fully autonomous AI system placed in a featured Kaggle competition with prize money.^[2]

---

Product Overview

Autoscience operates as a lab with one commercial product. Carl is the research engine — hypothesis generation, experimentation, and paper writing — whose output the company has pushed through academic peer review as its public benchmark.^[3] Mira packages that engine for companies running production ML: it "drops into your stack, reads your codebase, and starts shipping verified improvements within a week," continuously parsing new papers and recommending or implementing the techniques that fit.^[6][8]

The pitch targets the research-to-production lag. CEO Eliot Cowan: companies take "months, or even more than a year, to incorporate the latest research into their models," which he frames as "six- or seven-figure differences to their bottom line" — against an arXiv firehose of more than 2,000 ML papers per week.^[2] The product page asks: "What will you do with 100 ML researchers?"^[6]

Key Capabilities

Product Surfaces

---

Technical Architecture

Autoscience discloses little architecture. Mira's described loop is: ingest the customer codebase, continuously read and parse new research papers, match relevant techniques to the repositories, then implement and verify improvements — with deployment "within a week" of dropping into a stack.^[6][8] The underlying models, infrastructure, and verification methodology are not publicly documented; the strongest public evidence of capability is external — peer-review acceptances (later withdrawn) and the Kaggle Santa 2025 silver medal against 3,300 teams.^[7][2]

Key Technical Details

---

Strengths

• The strongest external validation in the category — double-blind workshop acceptances at ICLR 2025 and a Kaggle featured-competition silver medal against 3,300 teams are independent tests, not vendor benchmarks.^[3][2]
• A commercial wedge, not just a demo — Mira sells deployed model improvements into production codebases, "the output of a full ML research division without the headcount," where most autoresearch systems stop at papers.^[2][6]
• Well-capitalized for the niche — $14M seed led by General Catalyst with strategic backers (Toyota Ventures, Perplexity Fund) signals investor conviction in ML-research automation specifically.^[2]
• Demonstrated progression — from short workshop papers (March 2025) to a reported full-length peer-reviewed paper, suggesting the research engine is improving rather than a one-off stunt.^[3][4]

---

Cautions

• The headline milestone is tainted — the ICLR papers were submitted without organizers' knowledge, condemned by academics as co-opting peer review for publicity, and withdrawn; "first AI to pass peer review" carries an asterisk the marketing omits.^[7]
• Mira is early access with no public pricing, docs, or named customers — every capability claim ("verified improvements within a week," novel architectures) is vendor-stated and independently unverified.^[6][8]
• Opaque company — only CEO Eliot Cowan is publicly identified; team size, founding date, and architecture are undisclosed, thin diligence material for buyers handing an agent their production ML code.^[2]
• The Kaggle and peer-review wins measure research output, not the product — there is no public evidence yet that Mira's deployed improvements deliver the claimed six-to-seven-figure impact.^[2]

---

What Developers Say

Community sentiment is dominated by the March 2025 peer-review controversy, and it skews sharply critical; no substantive independent praise for the product itself was found as of June 2026 — Mira is too new and too closed for hands-on developer discussion, which is itself a data point.^[7][8]

"All these AI scientist papers are using peer-reviewed venues as their human evals, but no one consented to providing this free labor" — Prithviraj Ammanabrolu, assistant professor of computer science, UC San Diego^[7]

"Submitting AI papers to a venue without contacting the [reviewers] is bad" — Ashwinee Panda, postdoctoral fellow, University of Maryland, citing a "lack of respect for human reviewers' time"^[7]

"Evals [should be] done by researchers fully compensated for their time… Academia is not there to outsource free [AI] evals" — Alexander Doria, co-founder of Pleias^[7]

Autoscience responded by withdrawing the accepted papers, acknowledging organizers had not had time to set standards for autonomous-AI submissions, and proposing a dedicated venue for AI-generated research.^[7][3]

---

Pricing & Licensing

Pricing is not publicly listed as of June 2026.^[6]

Licensing model: Proprietary managed service; no open-source components.^[1]

Hidden costs: Unknown — with no public pricing or contract terms, total cost (and the engineering time to review agent-generated changes to production models) cannot be estimated in advance.^[8]

---

Competitive Positioning

Direct Competitors

When to Choose Autoscience Over Alternatives

• Choose Autoscience when: you run production ML models, the research-to-production lag costs you real money, and you want a managed agent that implements (not just summarizes) new techniques into your codebase.
• Choose AI Scientist when: you want an open-source, auditable autoresearch system with peer-reviewed methodology you can run and modify yourself.
• Choose Kosmos when: your problem is scientific discovery beyond ML model optimization.
• Choose pi-autoresearch when: you want self-run optimization loops over any measurable metric without handing code to a closed vendor.

---

Ideal Customer Profile

Best fit:

• Companies whose ML models drive revenue (recommendations, forecasting, ranking) and who can't staff a research team to track 2,000+ weekly arXiv papers^[2]
• Teams comfortable being early-access design partners with a seed-stage vendor inside their model code
• Organizations that can measure model improvements rigorously enough to verify a vendor's "verified improvements" claim themselves

Poor fit:

• Buyers needing public pricing, documentation, references, or a self-serve trial before committing
• Security- or IP-sensitive organizations unwilling to give a closed agent access to production ML codebases
• Academic groups — the peer-review controversy makes Carl-style submission workflows institutionally radioactive for now^[7]

---

Viability Assessment

The capability trajectory is real and fast: external, adversarial tests (peer review, Kaggle) twelve months apart show a system that improved materially, and General Catalyst leading a $14M seed off that record is meaningful validation.^[2] The open question is trust — the company's signature achievement came via a process academics consider abusive, and its commercial product asks customers for deep codebase access with nothing publicly verifiable behind it yet.^[7][8]

---

Bottom Line

Autoscience has the most credible raw capability story in autoresearch — its agents passed double-blind peer review and medaled in a featured Kaggle competition, tests no marketing department can fake — and the clearest business model: Mira turns the arXiv firehose into deployed improvements in customer models. But the peer-review milestone was won by ambushing volunteer reviewers and then withdrawn, the product is closed early access with no pricing or references, and the only community voice on record is academic condemnation.

Recommended for: ML-driven companies with money-losing research lag who want to be early design partners and can independently verify model improvements.

Not recommended for: Buyers needing transparency, references, or self-serve evaluation; IP-sensitive organizations; anyone in academia.

Outlook: Watch for named Mira customers with verifiable results, public pricing, and whether the proposed venue for AI-generated research materializes — converting a contested stunt into a legitimate channel would resolve the company's central reputational liability.^[7]

---

Research by Ry Walker Research • methodology