Tilde Interview Guide (2026): Inside the Khosla-Backed AI Interpretability Lab

First, make sure you have the right company. The Tilde that tech candidates are searching for in 2026 is Tilde Research (tilderesearch.com), the Palo Alto AI lab working on mechanistic interpretability, new architectures, and pretraining science. It is unrelated to tilde.ai, a European language-technology firm, and unrelated to the Portland software consultancy whose old reviews show up on Glassdoor under the same name. Tilde Research launched publicly in November 2024 and closed an $8M seed round led by Khosla Ventures in mid-2025, with Soma Capital and angels reported to include researchers at OpenAI and Anthropic. The founding team came out of Stanford; co-founder Dhruv Pai is the chief scientist. The company builds interpreter models and control tooling that let customers dissect model behavior and edit failure modes, and it has published open work such as Activault (a low-cost activation-storage engine) and Sieve (a case study where sparse autoencoders beat steering and prompting baselines on a code-generation task). Headcount was in the single digits per PitchBook as of early 2026, and that fact shapes everything about how they hire: you will be evaluated by founders, quickly, on whether you can do the actual work.

Process

Public data on Tilde’s loop is thin. There are no Glassdoor or Blind entries for this company (again, the Glassdoor “Tilde” reviews belong to a different firm), so what follows is the likely shape based on their job postings, their stage, and how comparable seed-stage AI labs run hiring. Applications go through Ashby or directly to [email protected], and the open roles as of mid-2026 are ML Engineer, ML Researcher, and Kernel Engineer, each offered as both internship and full-time, all listed in Palo Alto. Expect an intro conversation with a founder (there is no recruiting team at this size), then one or two technical rounds that mix live coding or kernel work with discussion of your past projects, then a research-taste conversation about interpretability and where the field is going. Small labs like this often add a short paid work trial or take-home project before an offer; candidates at peer labs mention this pattern often, though we cannot confirm Tilde uses one. The whole cycle should run one to three weeks, and decisions tend to be fast in both directions.

What they actually ask

• Walk me through something you built. Every posting asks for demonstrated ability through open-source contributions or technical writing, so expect a long, detailed walkthrough of your best public work: design choices, failure modes, what you would redo.
• ML infrastructure design. The ML Engineer role centers on designing end-to-end training and evaluation pipelines in PyTorch or JAX, scaling experiment throughput, and keeping distributed systems healthy. Expect questions shaped like system design but aimed at GPU clusters: activation storage, checkpointing, data loading, multi-node failure recovery.
• Kernel and GPU questions. For the Kernel Engineer track: write or optimize a Triton kernel, reason about memory hierarchy and occupancy, explain why a given kernel is slow. Postings name Triton, ThunderKittens, and TileLang, plus CUDA familiarity and GPU architecture knowledge.
• Interpretability literature. Sparse autoencoders, steering vectors, circuits-style analysis, and when each beats plain prompting. Their Sieve post argues SAEs can win on real tasks; expect to discuss that claim critically.
• Research ideation. The ML Researcher posting asks for depth in a subfield (architectures, optimizers, RL, learning dynamics) and the ability to propose experiments. Come with two or three concrete research ideas you could defend.
• Practical coding. Likely closer to real engineering than puzzle trivia, but a warm-up refresher on coding patterns is cheap insurance since algorithmic thinking is called out explicitly in the kernel posting.

Levels and comp (2026)

There is no Tilde data on levels.fyi, and the Ashby postings do not publish salary bands, so any specific number you see quoted elsewhere is guesswork. At a seed-stage lab with $8M raised, expect a flat structure with no formal levels: comp is negotiated per person. Based on what peer seed-stage AI research labs in the Bay Area were paying in 2025 and 2026, a reasonable expectation for engineers is roughly $150K to $250K base with equity in the 0.1% to 1%+ range depending on seniority, and kernel or ML-infrastructure specialists can command the top of that band or above given how contested that talent pool is right now. Internships at comparable labs pay roughly $8K to $12K per month. Treat all of that as market context, no more. In the actual conversation, ask directly for the base and equity split, the option strike price, the preferred price from the Khosla round, and the current fully diluted share count, since a percentage without those numbers is not interpretable, which is a fitting problem for this company.

Prep priorities

Prioritize in this order. First, read everything Tilde has published: the Sieve and Activault posts, their launch material, and their GitHub (tilde-research). Interviewers at labs this small expect you to know their work, and the fastest way to fail is arriving unfamiliar with it. Second, build or polish one piece of public evidence matched to your target role: a Triton kernel with benchmarks, a training-infrastructure tool, or a small interpretability replication (an SAE trained on a small model is a weekend project and directly on-topic). Third, drill GPU-flavored systems thinking, since infrastructure discussion is likely the longest technical segment for the engineering roles. Fourth, read the core interpretability papers from Anthropic and the open-source SAE literature so the research conversation has depth. Finally, prepare crisp stories about ownership, speed, and why you want a tiny in-person lab over a big-lab job; the behavioral portion will be informal but it decides whether founders want to sit next to you every day.

Frequently Asked Questions

Is this the same company as tilde.ai? No. tilde.ai is a long-established European language-technology and localization company based in Latvia. The AI startup drawing interview interest is Tilde Research at tilderesearch.com, founded in 2024 in Palo Alto. Confirm which one your recruiter or job posting belongs to before you prep.

Do I need a PhD to get hired? Reports suggest no. The lab has described itself as open to junior candidates, offers internships across all three tracks, and its postings emphasize demonstrated work (open source, technical writing, shipped systems) over credentials. A strong public artifact beats a thin publication record here.

Is remote work possible? Assume not. All three postings list Palo Alto, and labs at this headcount almost always default to in-person. It costs nothing to ask, but plan your search around being on-site in the Bay Area.

How should I choose between the ML Engineer and ML Researcher tracks? Pick based on your evidence, since the interviews will follow it. If your strongest work is pipelines, throughput, and distributed training, apply as an engineer; if it is papers, experiments, and learning-dynamics intuition, apply as a researcher. At a team this size the roles blur in practice, and candidates mention that small labs will happily re-route you mid-process if your background fits the other track better.