Find your next role

Discretionary grants for independent researchers, small teams, and short-term projects working on AI safety, alignment, technical governance, and adjacent long-term-risk topics. Typical award size $5K–$200K; rolling applications reviewed quarterly. No institutional affiliation required — designed for independent researchers, sabbatical work, and exploratory projects that don't fit traditional academic or industry funding cycles. Application is a short structured proposal; turnaround is typically 4–8 weeks.

Vetted freelance network placing AI / ML engineers into project-based engagements with product teams from early-stage startups to Fortune 500s. Engagement model is hourly or fixed-fee; rates set by the talent, not the platform. Admission requires passing a multi-stage technical screen + reference check (advertised as top 3% acceptance). Strong fit for engineers with shipped production AI work who want optionality over full-time commitment. Toptal handles client sourcing, contracts, and invoicing; talent retains direct relationships post-engagement.

Cursor is the AI-native code editor that grew from a few thousand users to one of the highest-revenue developer-tools companies in under two years. Engineering hires are extremely selective — small team, very high product velocity, in-person only. What you'd actually work on: editor performance (it's a fork of VS Code with very heavy AI integration), agent orchestration (Composer, the multi-file edit flow), the diff/merge UX that makes AI edits acceptable to professional engineers, or the inference-serving layer that handles trillions of tokens monthly. Honest fit signals: — You ship fast and don't need a roadmap document handed to you. Cursor engineers are expected to identify their own highest-leverage work and defend the choice to the leadership team weekly. — You're a power user of AI coding tools (ideally Cursor itself) and can articulate where the product is good vs where it falls down. — SF-resident. Cursor is remote-hostile on principle, not as policy theater. Compensation reflects that. What's not a fit: anyone who wants stable hours, anyone who isn't already fluent with React + TypeScript + low-level perf debugging, or anyone looking to learn AI on the job — the bar assumes you already use it daily.

Sierra builds conversational AI agents for enterprise customer experience — think Sonos, ADT, WeightWatchers running their support flows on Sierra's agent platform. Co-founded by Bret Taylor (ex-Salesforce co-CEO, OpenAI board chair) and Clay Bavor (ex-Google Labs). What you'd work on: the agent runtime (deterministic skills + LLM reasoning + tool use), evaluation infrastructure (every agent has to pass thousands of test scenarios before production), customer-specific fine-tuning pipelines, or the developer experience for the people who build agents on the platform. Honest fit signals: — Strong systems engineering background — Sierra agents handle real customer money/data, so reliability and observability matter as much as capability. — You've thought about LLM evals seriously. Sierra's evaluation framework is arguably their biggest moat; engineers who join expecting "just prompt the model" find the real work surprising. — You're motivated by the enterprise-AI problem space (not consumer). The customers are big, demanding, and pay seven figures. What's not a fit: anyone allergic to enterprise sales cycles influencing engineering priorities, anyone who wants pure research, or anyone who needs hard remote.

MATS is the canonical entry point into AI alignment research — a 10-week in-person fellowship in Berkeley where you work directly with a senior mentor (typically from Anthropic, OpenAI, DeepMind, Redwood Research, ARC, or independent senior researchers) on a concrete alignment research project. How it actually works: you apply to one or more mentors in the open call, complete a short technical assessment for each, and matched scholars relocate to Berkeley for the cohort. Each scholar produces a research artifact (paper, agenda, dataset, or demonstration) by the end of the program. The strongest MATS alumni are recruited directly into the major alignment labs. Honest fit signals: — Some prior technical depth — typically a CS/ML/math undergrad or equivalent, plus demonstrable curiosity about alignment specifically (read at least the major papers on the MATS reading list before applying). — You're willing to commit to in-person Berkeley for 10 weeks. MATS treats the cohort dynamic as core, not optional. — You're not just job-seeking — selection is competitive and the mentors invest serious time. People who treat it as a credential burnout faster than people genuinely motivated by the research questions. Strong outcomes from MATS alumni: full-time alignment roles at major labs, independent research grants, founding small alignment-focused research orgs.

Astera funds scientists and engineers to pursue ambitious "focused research organization" (FRO) ideas — projects that fall between traditional academic grants and venture-backed startups. Previous residents have spun out FROs on AI evals, biorisk monitoring, metascience, and large-scale dataset construction. The residency is open-ended in topic but selective in shape: Astera funds people who already have a clear, ambitious research project and need 12–24 months of runway + institutional support to prove it works, then either spin it out as an FRO or wrap it as a research artifact. Honest fit signals: — You already have a specific, defensible project idea — not "I want to explore X." Astera's residency interviews are essentially venture pitches for non-profit research bets. — You can articulate the impact mechanism (what changes if this works, who benefits, why won't existing institutions do it). — You're senior enough to lead the project independently. The residency is not a postdoc-style learning fellowship. What's not a fit: open-ended PhD-style research, junior researchers who need close mentorship, or commercial product ideas (those want a VC, not Astera).

The original startup accelerator. Three months in-person in San Francisco, the YC partner network and Bookface community for the life of the company, Demo Day in front of ~1,500 investors, and a network of 4,000+ alumni founders. For AI-skilled candidates specifically: YC's recent batches have been ~70%+ AI companies. The bar to get in has risen accordingly — having shipped product, talked to real users, and shown some indication of demand is increasingly the floor, not the ceiling. Honest fit signals: — You have a co-founder (or are open to YC's co-founder matching). Solo founders get in but face higher scrutiny. — You can articulate the wedge — what's the specific 1-foot-wide doorway you're going through, not the 50-foot vision you'll execute against later. — You're willing to relocate to SF for the batch. YC has experimented with remote/hybrid; the in-person batches consistently produce better outcomes by every metric YC tracks. What's not a fit: lifestyle businesses, anyone unwilling to take outside capital, or anyone who's still in the "exploring ideas" phase rather than committed to building one specific thing for the next 5–10 years.

OpenAI's research internship places you on a research team — pre- training, post-training, alignment, reasoning, robotics, or applied research — for 12 weeks over summer, working on a defined research project alongside a mentor who's a member of technical staff. What the role actually involves: a single concentrated research project (not multiple "exploratory" mini-projects), real compute budget, real codebase contributions, and a formal write-up at the end. Strong interns are offered full-time conversion at the end of the summer. Honest fit signals: — Currently enrolled in a PhD program (the typical profile) OR an exceptional undergraduate/masters student with a strong publication or open-source track record in ML. — You can write fluent PyTorch and have already done multi-GPU / multi-node training (or are a fast learner the mentor judges credible). — You're comfortable with the public scrutiny that comes with the OpenAI brand. The research culture is genuinely high-stakes. What's not a fit: anyone using the internship as a "try out AI" — the bar is calibrated to people who already know they want to do frontier ML research and have early evidence of capability.

Surge AI hires domain experts as contract evaluators for frontier-lab training pipelines — your job is to design, write, or judge outputs against rubrics in your specialty (law, medicine, finance, academic research, software engineering, creative writing, math, etc.). The labs use the resulting data for RLHF, evals, and red-team curricula. How the engagement works: paid hourly, project-based, flexible hours (typically 10–30hrs/week while a project runs), remote, projects last weeks to months. Specialist rates apply if you have credentialed expertise (e.g. licensed attorney, MD, working software engineer at a senior level) — Surge has a known-good rate card you'll see during onboarding. Honest fit signals: — Deep domain expertise that the labs can't crowd-source. Generalist applicants get routed to lower-rate tasks and may not be matched at all in tight-domain projects. — You can write clearly and follow detailed rubrics. The work IS judgment calls, but they're judgment calls within tight definitions. — You're comfortable with NDA-bound contract work (you won't be allowed to talk publicly about specific projects or labs you work with). What's not a fit: anyone looking for full-time work (Surge is genuinely contract-only), anyone without specialist credentials hoping to make top rates, or anyone uncomfortable with their judgments being used to train commercial AI systems.

You'll work directly on training and deploying frontier large language models — pre-training, fine-tuning, RLHF, evals, interpretability, or alignment research, depending on where your strengths fit. What the role actually involves: writing production training code at scale (PyTorch, large distributed clusters), running rigorous experiments with clear hypotheses, and communicating results to research leadership. Less about novel academic ideas; more about disciplined execution against an existing research agenda. Honest fit signals: — You've shipped ML systems that ran in production, not just notebook prototypes. Multi-node training experience matters here. — You're comfortable with the safety-first ethos: this is the team that wrote Constitutional AI and runs the Frontier Red Team. If you're primarily motivated by capability scaling, the culture friction will be real. — Strong written communication. Anthropic's engineering culture is doc-driven; you'll write design docs, postmortems, and research notes more than you'd expect. What's not a fit: pure researchers without engineering taste, engineers without ML systems experience, or anyone who wants a faster product iteration loop than a safety lab will give you.