What is Mirendil's position vs Sakana AI?

Both Mirendil and Sakana AI build AI systems aimed at accelerating AI research itself, but the approach and geography differ. Sakana AI is a Tokyo-based AI lab founded in 2023 by ex-Google Brain researchers David Ha and Llion Jones (also a Transformer co-author). Sakana's focus is evolutionary and bio-inspired AI methods — model merging, evolutionary architecture search, and self-improving systems. Notable Sakana projects include the AI Scientist (an end-to-end research agent) and Continuous Thought Machines. Mirendil, which closed a $200M seed at $1B valuation on June 25, 2026 with a16z and Kleiner co-leading, focuses on AI systems that help researchers iterate faster on better AI in specialized domains (biology, chemistry, medicine, materials, robotics). The teams overlap in pedigree (both have ex-Google Brain founders) but the philosophies differ — Sakana bets on biological/evolutionary metaphors; Mirendil bets on AI-as-research-partner across scientific domains. Mirendil is also dramatically better-funded; Sakana's last reported round was much smaller.

How does Mirendil compare to Goodfire?

Mirendil and Goodfire address adjacent but different research bottlenecks. Goodfire focuses on mechanistic interpretability — building tools to understand how neural networks compute, with the goal of editing model behavior at the circuit level rather than via fine-tuning or prompting. Goodfire's Ember platform exposes neuron-level features for editing and analysis. Mirendil's stated focus is broader: AI systems that accelerate AI research velocity in scientific domains, including but not limited to interpretability. The two could be complementary — Mirendil's research-acceleration AI could use Goodfire-style interpretability tools as one component, and Goodfire's interpretability platform could itself be accelerated by Mirendil-style research AI. For now, both are highly-funded bets on different layers of the AI-for-AI-research stack — Goodfire on understanding how models work, Mirendil on helping researchers build better ones.

Will frontier labs (Anthropic, OpenAI, Google DeepMind) build Mirendil's product internally?

Yes, and this is the most direct competitive risk to Mirendil's thesis. All four major frontier labs have internal research-acceleration tools — code agents for ML research, experiment management systems, hypothesis-generation tools — typically with privileged access to internal model APIs, internal datasets, and internal compute. Anthropic and OpenAI in particular have publicly discussed using their own models to accelerate alignment and capabilities research. Mirendil's defense rests on three things: (1) cross-lab portability — Mirendil tools work across Anthropic, OpenAI, Google, xAI, and emerging labs, while internal tools at each lab are typically lab-specific; (2) external research community — Mirendil can serve universities, startups, scientific institutions, and corporate research labs that don't have frontier-lab-scale internal tooling; (3) talent advantage — Mirendil's 20-person team drawn from all four labs has visibility into what's missing across the entire ecosystem. Whether these defenses hold against well-resourced internal efforts is the central execution question.

What's the difference between AI-for-AI-research and AutoML?

AutoML (Automated Machine Learning) refers to tools that automate model selection, hyperparameter tuning, feature engineering, and architecture search — typically for traditional ML workloads (tabular data, classical tasks). Examples include Google AutoML, AutoGluon, and H2O. AI-for-AI-research, the Mirendil category, is broader and more ambitious: it includes hypothesis generation, literature review, experiment design, code implementation, result analysis, paper writing, and ultimately autonomous research direction in scientific domains. AutoML automates the mechanics of training a model; AI-for-AI-research aims to automate the cognitive work of being a scientist. The two are conceptually related — AutoML is a subset of the broader research automation thesis — but the ambitions and target users differ substantially. AutoML tools serve practitioners; AI-for-AI-research tools target researchers themselves.

Quick Answer

Mirendil vs Sakana vs Goodfire: AI Research Labs 2026

Published: June 26, 2026

Mirendil vs Sakana vs Goodfire: AI Research Labs 2026

Mirendil closed a $200M seed at $1B valuation on June 25, 2026 to build AI systems that accelerate AI research itself. It joins a small but growing category of companies betting that research velocity — not raw compute — is the new bottleneck in AI capability gain. How does Mirendil compare to Sakana AI, Goodfire, and the frontier labs’ internal efforts? Short answer: Mirendil is the best-funded and broadest-scope bet in the category, but the competitive set is sophisticated and the thesis is contestable.

Last verified: June 26, 2026.

TL;DR

Mirendil: $200M seed at $1B valuation, June 25, 2026; team of 20 from Anthropic/OpenAI/DeepMind/xAI; broad AI-for-AI-research thesis
Sakana AI: Tokyo-based; evolutionary and bio-inspired AI methods; AI Scientist project; ex-Google Brain founders
Goodfire: mechanistic interpretability platform (Ember); neuron-level model editing
Frontier labs (internal): Anthropic, OpenAI, DeepMind, xAI all have internal research-acceleration tools
AutoML platforms: Google AutoML, H2O, AutoGluon — narrower scope, traditional ML focus
Bet: if research velocity becomes the dominant bottleneck, this category produces a category-defining outcome; if pre-training scale keeps mattering most, it stays niche

The category framing

These companies share a thesis: AI capability is now bottlenecked by research velocity, not compute. They differ in which part of research velocity they target.

Layer	Companies
Understanding models	Goodfire (mechanistic interpretability)
Building better algorithms	Sakana (evolutionary methods, model merging, self-improvement)
Accelerating research process	Mirendil (AI-as-research-partner across domains)
Automating ML practice	Google AutoML, AutoGluon, H2O (traditional ML)
Internal lab tools	Anthropic, OpenAI, DeepMind, xAI (proprietary)

Each layer is a different theory of what’s hard about doing AI research.

Lab-by-lab

Mirendil

Dimension	Detail
Founded	2024-2025 (public emergence June 2026)
Funding	$200M seed at $1B valuation (a16z + Kleiner Perkins co-led, NVIDIA participated)
Team	20 researchers from Anthropic, OpenAI, DeepMind, xAI
Founders	Behnam Neyshabur (deep learning theory) + Harsh Mehta
Focus	AI systems that accelerate AI research in specialized domains
Target users	AI/ML researchers, scientific institutions, well-funded labs
Differentiator	Broad cross-lab pedigree; aggressive valuation; explicit “AI for AI research” framing
Stage	Pre-product

Mirendil is the broadest and best-funded bet in the category. Its public emergence in the same week as the Sail Research and General Intuition rounds, plus the senior-talent departures from Google DeepMind, places it at the center of the June 2026 narrative on the next leg of AI investment.

Sakana AI

Dimension	Detail
Founded	2023
Headquarters	Tokyo, Japan
Founders	David Ha + Llion Jones (Transformer co-author) — both ex-Google Brain
Focus	Evolutionary and bio-inspired AI methods
Notable projects	AI Scientist (autonomous research agent), Continuous Thought Machines, model merging at scale
Differentiator	Evolutionary methods as core thesis; Japanese / Asia-Pacific positioning; strong public-research output
Stage	Multiple products, ongoing research

Sakana is the most public-research-active of the category. The AI Scientist project is the most direct prior art for “AI that does AI research” — Sakana published the AI Scientist in 2024 and has iterated since. Sakana’s bet on evolutionary methods is unusual and could be either visionary (if evolutionary self-improvement becomes a key capability) or niche (if gradient-based methods continue to dominate).

Goodfire

Dimension	Detail
Founded	~2024
Funding	Multiple rounds; latest reported $50M Series A
Focus	Mechanistic interpretability platform
Notable products	Ember (neuron-level feature exposure and editing)
Differentiator	Single-domain depth (interpretability); platform business model
Stage	Live product (Ember)

Goodfire is the cleanest “depth over breadth” bet. Mechanistic interpretability is a specific research area within AI safety; Goodfire’s bet is that interpretability tools will become essential infrastructure for any lab serious about understanding their own models. This complements rather than competes with Mirendil’s broader thesis.

Frontier labs (internal efforts)

All four frontier labs have internal AI-for-AI-research tools, though most details are private.

Anthropic: internal code agents using Claude for ML research, alignment research tools, internal Claude variants for specific research tasks
OpenAI: publicly discussed using internal agents (Codex variants) to accelerate ML research; mention of AI scientist projects in technical reports
Google DeepMind: Project Astra and related research tools; significant internal investment in AI-assisted research
xAI: less public about internal tooling but Elon Musk has publicly discussed using Grok internally for research acceleration

These internal efforts are Mirendil’s most direct competition. The four frontier labs have privileged access to their own models, internal compute, and proprietary datasets — advantages an external company cannot match.

AutoML platforms (adjacent but different category)

Platform	Focus	Differentiator
Google Cloud AutoML	Traditional ML automation	Vertex AI integration
AutoGluon	Open-source AutoML	Strong tabular performance
H2O Driverless AI	Enterprise AutoML	Long-established enterprise customers
DataRobot	Enterprise AutoML	Vertical solutions, governance

AutoML automates the mechanics of training a model on a given dataset. Mirendil’s thesis is broader: automate the cognitive work of being a researcher (hypothesis generation, experiment design, result analysis, paper writing). AutoML is a subset of the broader AI-for-AI-research thesis but is mature and largely commoditized.

Where Mirendil is most differentiated

Breadth + funding

No other independent company in the category has Mirendil’s combined breadth of mission (AI-for-AI-research across scientific domains, not narrow on interpretability or evolutionary methods) and funding level ($200M is roughly 4-10x what the next-most-funded independent peer has raised).

Cross-lab team

The 20 founding researchers drawn from Anthropic, OpenAI, DeepMind, and xAI is itself a research asset. Each lab has different proprietary methods; assembling a team with visibility into all of them is unusual. Sakana, Goodfire, and others have strong teams but less cross-lab breadth.

Investor signal

a16z + Kleiner Perkins co-lead with NVIDIA participation is the strongest possible signal that the venture market sees this as a category-defining bet. By contrast, Sakana has raised meaningful but smaller amounts; Goodfire has raised at less aggressive valuations.

Where the competitive risk is highest

Frontier labs build it themselves

The biggest risk. If Anthropic, OpenAI, or DeepMind ship excellent internal research-acceleration tools and choose not to externalize them (because the productivity gain matters more than the revenue), Mirendil’s market shrinks substantially.

Sakana’s research lead

Sakana has been publishing AI-for-AI-research papers since 2023. The AI Scientist work is the most direct prior art. Mirendil enters with more money but has to demonstrate research output to catch up on the public-research front.

Specialization wins

In several historical analogous waves, narrow specialists beat broad generalists. Goodfire’s depth on interpretability could outcompete Mirendil’s breadth on the same problem space. Sakana’s depth on evolutionary methods could win that subset.

Talent retention

In a 2026 talent market where senior researchers regularly switch labs for $100M+ packages, retention is the most operational risk. Any one of Mirendil’s 20 founding researchers could be poached for a frontier-lab package or to start a competing company.

How to think about this as a builder

For most AI builders, none of these companies are directly relevant in 2026. The deliverables are upstream (better models, better research velocity); your application benefits indirectly when better models become available.

For ML researchers and AI/ML engineers at non-frontier labs: watch for first product releases from Mirendil (expected 2027) and Sakana (continuing to ship). These could become important workflow tools.

For investors and category-watchers: Mirendil is the cleanest bet on AI-for-AI-research as a venture-scale category. Its trajectory signals whether the category produces a single dominant company, a fragmented set of specialists (Sakana on evolution, Goodfire on interpretability), or gets absorbed into the frontier labs’ internal efforts.

Bottom line

Mirendil is the most ambitious and best-funded bet on a contested thesis: that AI research velocity is the new bottleneck, and AI-for-AI-research becomes the highest-leverage layer in the stack. Sakana AI is the most public-research-active alternative with a distinctive evolutionary-methods angle. Goodfire is the depth-on-interpretability complement. Frontier labs are the largest competitive risk because they can build internal versions of all of these.

The category will likely produce a mix: at least one large independent winner (potentially Mirendil if execution works), at least one narrow specialist outcome (potentially Goodfire), and substantial internal capability inside frontier labs that never becomes a public product. Which of those Mirendil ends up as will be visible by 2028.