Tilde

What Tilde does

Tilde Research is a venture-backed AI research lab founded in 2024, focused on mechanistic interpretability, novel neural network architectures, and pretraining science. The company develops foundational understanding of model computation to advance the frontier of AI capabilities and safety. Its core product portfolio consists of open-source research tools, including Sieve (an SAE-based framework for conditional feature steering in code generation with greater than 99.9% precision), Activault (an S3-based activation data engine that reduces training costs for interpreter models by 4-8x), and MoMoE (a memory-optimized Mixture-of-Experts Triton kernel outperforming NVIDIA's TEGrouped and ScatterMoE). Additional technical releases span the Aurora optimizer (100x data efficiency at 1.1B parameter scale), Parallax local linear attention, Wall Attention (a RoPE replacement), and Nitrobrew distillation (1.5-3x faster throughput).

The company operates a horizontal go-to-market targeting the AI research community, distributing all outputs freely through GitHub, HuggingFace, and a technical blog. There is no commercial pricing model, API, or paid product; the business is sustained by venture capital, with an $8M round closed in July 2025 led by Khosla Ventures, alongside Bain Capital Ventures, Forgepoint Capital, Soma Capital, and Timeless. The team includes researchers from OpenAI and Anthropic, with academic collaboration from Yufei Zhao, and infrastructure support from Nebius AI. Customer segments are exclusively research-oriented (academic researchers, independent researchers, small research groups working on interpretability and AI safety), with no disclosed paying customers or enterprise contracts.

Name

Tilde

Legal name

Tilde Research

Website

https://tilderesearch.com

Company type

Private

Founded year

2016

Operating status

Operating

Headcount range

11–50 employees

Short description

Tilde Research is a venture-backed AI research lab developing open-source tools for mechanistic interpretability, novel neural architectures (optimizers, attention, MoE), and pretraining science, serving the AI research community with no commercial product.

Ownership category

akta.pro rank

Product category

AI Research Tools

NAICS

Scientific Research and Development Services (5417), Research and Development in the Physical, Engineering, and Life Sciences (except Nanotechnology and Biotechnology) (541715)

SIC

Services-Commercial Physical & Biological Research (8731), Services-Engineering, Accounting, Research, Management (8700)

akta.pro primary industry

Synthetic Data & Data Augmentation for Foundation Models (HDAAACAK)

Keywords

• Mechanistic interpretability
• Open-source AI research
• Sparse autoencoders
• Neural network optimization
• AI safety research

Headquarters

HQ city

San Francisco

HQ country

United States

HQ region

North America

Markets served

GTM type

B2B

Offering type

Software

Cost components

Personnel, Technology or R&D, Infrastructure, Operations

Distribution channels1 record

Marketing channels5 records

Core offering

Tilde Research develops and open-sources applied interpretability research artifacts and novel neural network components for AI safety and frontier model development. Its deliverables include mechanistic interpretability frameworks (Sieve), activation data infrastructure (Activault), high-performance training kernels (MoMoE, NSA), optimizers (Aurora), attention mechanisms (Parallax, Wall Attention), and distillation techniques (Nitrobrew), all released as research tools rather than commercial products.

Product overview

Tilde Research is a moonshot AI lab focused on mechanistic interpretability, new architectures, and pretraining science. The company develops open-source research tools and publishes findings rather than offering a commercial product. Core offerings include: Sieve (SAE-based model control framework for code generation), Activault (S3-based activation storage engine), and MoMoE (Triton kernel for MoE optimization). Additional technical releases include Aurora optimizer, Parallax/Wall Attention (attention mechanism improvements), Nitrobrew distillation, and NSA sparse attention implementation. All products are released as open-source code and models.

Differentiator

Problem solved

Functional benefit

Products and services

• Sieve Sieve is an applied interpretability framework that uses sparse autoencoders to perform conditional feature steering on large language models, enabling automatic interventions such as suppressing regex usage in generated Python fuzz tests while preserving code functionality, with greater than 99.9% precision. It is targeted at mechanistic interpretability researchers and ML engineers working on LLM behavior control.
• Activault Activault is an activation data engine designed to reduce the cost of training interpreter models by generating, caching, and streaming model activations through S3-compatible object storage. It is aimed at researchers and ML engineers who need to collect and store large volumes of model activations for mechanistic interpretability work.
• NSA Implementation The NSA Implementation is a hardware-aware Triton-based implementation of Native Sparse Attention providing high-throughput training and evaluation with custom forward and backward kernels optimized for GPU execution. It targets researchers and engineers experimenting with sparse attention mechanisms for efficient long-context modeling.
• MoMoE MoMoE is a high-performance Triton kernel implementation for training and fine-tuning Mixture-of-Experts (MoE) models, offering significant speed and memory efficiency improvements over NVIDIA TEGrouped and ScatterMoE baselines. It is targeted at ML engineers running large-scale MoE training and inference workloads.
• Nitrobrew Nitrobrew is a fast, lossless distillation technique that exploits the fact that teacher logits are generated from a much lower-dimensional hidden state, enabling 1.5-3x faster end-to-end throughput for on-policy distillation. It targets researchers and engineers working on knowledge distillation for large language models.
• Aurora Aurora is a leverage-aware optimizer for training neural networks that addresses the neuron death problem in the Muon optimizer by reformulating update selection to jointly satisfy left semi-orthogonality and uniform row norms. It is aimed at ML researchers and engineers seeking more stable, data-efficient pretraining.
• Wall Attention Wall Attention is a data-dependent positional encoding that generalizes diagonal forget gates from linear RNNs to softmax attention, replacing RoPE with a mechanism that exhibits strong length generalization. It is intended for researchers and engineers designing attention layers for long-context models.
• Parallax Parallax is a parameterized local linear attention method that retains the local-linear principle while parametrizing the solve directly as a single additive correction on top of Softmax. It targets researchers and engineers exploring alternatives to standard softmax attention.

Quantifiable outcome

• Aurora optimizer achieves 100x data efficiency on open-source internet data at 1.1B parameter scale
• +4 more outcomes

Segments1 record

Ideal customer profiles2 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

AI capability11 records

Feature7 records

Partnerships

Four partnerships are on record, tiered key and supporting.

• Researchers at OpenAIkeyStrategic or Co-development PartnerActive researchers from OpenAI are associated with Tilde Research, bringing expertise from one of the leading AI research organizations to the lab's mission of mechanistic interpretability.
• Researchers at AnthropickeyStrategic or Co-development PartnerActive researchers from Anthropic are associated with Tilde Research, contributing deep expertise in AI safety and interpretability research to the lab's work.
• Yufei ZhaokeyStrategic or Co-development PartnerAcademic researcher providing expertise and collaboration with Tilde Research, contributing to the lab's research initiatives in mechanistic interpretability.
• Nebius AIsupportingOthersCloud infrastructure provider supporting Tilde Research's computational needs. Nebius AI provided assistance scaling Activault on their solution and contributed solutions architects for technical support.

Recent moves6 records

Expansion highlights5 records

Direct peers

• Goodfire: Independent mechanistic-interpretability startup building tooling and APIs for inspecting and steering frontier model internals. Closest direct peer to Tilde in mission, stage, and product surface area.
• Apollo Research: Independent AI safety research organization focused on interpretability and evaluation of advanced models. Shares Tilde's research-lab, AI-safety-first orientation and similar target audience of researchers and policy stakeholders.

Broad incumbents

• Anthropic: Frontier model lab that originated sparse-autoencoder-based interpretability (Towards Monosemanticity). Researchers from Anthropic collaborate with Tilde, and Anthropic runs the most comparable in-house interpretability program at scale.
• OpenAI: Frontier model lab with active mechanistic interpretability research (e.g., sparse autoencoders on GPT internals). Active OpenAI researchers are also associated with Tilde, making it both a strategic partner and a competitor.
• Google DeepMind: Large incumbent research lab with substantial mechanistic interpretability and architecture work. Represents the well-capitalized incumbent that Tilde's efficiency-focused research aims to leapfrog on cost.
• Mistral AI: Foundation-model lab that publishes open-source models and research, with comparable architectural and optimizer research interests. Represents a scaled peer that has successfully transitioned from research output to commercial deployments.

Emerging players

• Ndea: Early-stage research-oriented AI lab focused on mechanistic interpretability and model understanding. Comparable to Tilde in stage and thematic focus, though operating with a different technical approach.
• EleutherAI: Open-source AI research collective producing interpretability research, training infrastructure, and open models. Shares Tilde's open-source distribution model and research-community GTM.
• Sakana AI: Tokyo-based AI research lab that publishes novel architectures and training-method research rather than a commercial product. Closely comparable in 'research lab as company' structure and scientific ambition.

Others

• MATS (ML Alignment & Theory Scholars): Research fellowship program rather than a company, but functions as a talent pipeline and intellectual hub for mechanistic interpretability and AI safety research — directly comparable as a node in the same talent and ideas ecosystem Tilde operates in.

Market position

Strengths4 records

Weaknesses4 records

Competitive moat4 records

Key risks5 records

Key highlights6 records

Customer concentration

• Xx.com/tilderesearch
• GitHubgithub.com/tilde-research
• LinkedInlinkedin.com/company/tilderesearch

Revenue estimate

Valuation estimate

Number of profiles

Funding overview

Funding rounds1 record

Investors5 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Tilde