Landing a Job at a Frontier Lab

OpenAI

OpenAI makes ChatGPT and the GPT models. HQ in San Francisco, with hubs in New York, London, Dublin, Tokyo, and Singapore. Office-first and not meaningfully remote. Sam Altman has said publicly that the fully-remote experiment is over, and the few remote roles cluster in security, infrastructure, and deployment. Careers: openai.com/careers (the live board is on Ashby).

Anthropic

Anthropic is the safety-focused lab behind Claude. HQ in San Francisco, with offices in New York, Seattle, London, and Zürich. Hybrid and office-forward. Their careers page is refreshingly blunt about it: “Most staff are in the Bay Area and come to the office regularly. Some live further away and come in for one week a month.” Many roles are tagged “Remote-Friendly (Travel-Required).” Careers: anthropic.com/careers. If “Travel-Required” means going to the Bay Area for a week every month…yuck.

Google DeepMind

Google DeepMind is Alphabet’s unified AI lab, building Gemini. HQ in London (King’s Cross), with around ten offices including Mountain View, New York, Paris, Zurich, Toronto, Montreal, Bangalore, and Tokyo. Roles are anchored to a specific office city and sit under Google’s company-wide roughly three-day hybrid expectation, which has been tightening. Fully-remote roles are rare. Careers: deepmind.google/about/careers. If I’m being honest DeepMind is the most attractive place to me. I’d love to live in London, but I don’t think it fits my lifestyle very well. I’d have to move a family and (too many) pets halfway across the world. To have as much space as I do now I’d need to live at least an hour out by train (their office is directly at King’s Cross, which I like to think would help my finally fulfill my fantasy of being a Hogwart’s student).

Meta

Meta (Superintelligence Labs and FAIR) consolidated its AI work in 2025 under a new structure spanning frontier models, FAIR, applied research, and infrastructure. HQ in Menlo Park. Hybrid and in-office leaning, governed by Meta’s in-person policy of at least three days a week on site. Research roles may differ from the company default, so check the posting. Careers: metacareers.com, AI team. On a personal tangent: the ESM protein language models started at Meta’s FAIR before the team spun out to found EvolutionaryScale, which now builds ESM3 and ESMC. ESM is one of the projects that has my direct attention at the moment.

xAI

xAI builds Grok. HQ in Palo Alto, with major compute in Memphis. Mostly hybrid, with Palo Alto roles trending toward full in-office and a small number of remote US roles. Their careers page bot-blocks automated checks, so open it in a browser to confirm. Careers: x.ai/careers.

Microsoft AI

Microsoft AI (MAI), under Mustafa Suleyman, has moved from wrapping others’ models to training its own foundation models from scratch, and is a Frontier Model Forum founding member. HQ in Redmond, with MAI presence in Mountain View, New York, and London. MAI specifically has run a stricter in-office expectation than the company at large (around four days a week as of early 2026). Careers: microsoft.ai/careers. Microsoft Research is a separate, more open-publication org: microsoft.com/research/careers.

Amazon, Nova, etc

Amazon (AGI Lab and Nova) trains the Nova foundation models from scratch on its own Trainium and Inferentia silicon, and is on the Frontier Model Forum roster, though it has historically leaned toward price and performance rather than topping leaderboards. AGI labs in San Francisco and Boston; corporate HQ in Seattle. Amazon mandates five days a week in office company-wide. Careers: amazon.jobs (AGI team).

SSI

Safe Superintelligence Inc. (SSI) is Ilya Sutskever’s research-only company, with no public product by design. Offices in Palo Alto and Tel Aviv. Co-located in those two cities, with no remote language on the site. Careers: ssi.inc (board).

Thinking Machines Lab

Thinking Machines Lab is Mira Murati’s startup. It has shipped a fine-tuning API (Tinker) and research previews rather than a frontier flagship so far. Based in San Francisco, and every listed role is in San Francisco. Careers: thinkingmachines.ai (board). I keep seeing this listed as a Frontier Lab on other lists/sources, but I think, going by my original definition above, it doesn’t quite feel like the right classification.

Mistral AI is the Paris lab behind a mix of open-weight and commercial models plus Le Chat. HQ in Paris, with offices in London, Palo Alto, and New York. Hybrid-leaning and role and geography dependent, with relocation and visa support, and a handful of postings that allow remote. Careers: mistral.ai/careers (Lever board).

This is quite honestly all I could find atm, but that doesn’t mean there’s not more.

DeepSeek, the open-weight lab spun out of the quant fund High-Flyer, with offices in Hangzhou and Beijing. Careers run through MokaHR and recruiting platforms. Qwen / Alibaba’s Tongyi Lab in Hangzhou, hiring via talent.alibaba.com. Moonshot AI (Kimi) in Beijing, which unusually runs an English-capable Ashby board alongside its Chinese portal. Zhipu AI / Z.ai (GLM), a Tsinghua spinout in Beijing, now publicly traded, hiring through a Feishu portal. MiniMax in Shanghai, also now public, with an English careers page that routes applications through Feishu.

Cohere builds enterprise and sovereign foundation models (the Command family) out of Toronto, with offices in San Francisco, New York, Montreal, London, Paris, and Seoul. Remote-first culture. Careers: cohere.com/careers, ashby board. Some roles listed directly as Remote, others as Hybrid, though many list location directly.

Reka AI builds multimodal models (text, image, audio, video) across the SF Bay Area and Singapore, globally distributed and remote-friendly. Careers: jobs.ashbyhq.com/reka. Didn’t spend much time fact checking this one.

Liquid AI, an MIT CSAIL spin-off, builds efficient edge and on-device foundation models on non-transformer architectures. Based in the Boston area, with roles in SF, New York, and Tokyo. Primarily hybrid. Careers: jobs.ashbyhq.com/liquid-ai. From what I can tell despite listing some remote, you are largely expected to live either in SF or Boston.

Sakana AI in Tokyo uses nature-inspired and evolutionary methods, and is Japan’s most prominent AI lab. In-office in Tokyo for every listed role. Careers: sakana.ai/careers. I bet living in Tokyo would be pretty neat though!

Contextual AI builds enterprise RAG for regulated industries; its CEO co-authored the original RAG paper. Mountain View, with New York and London. Hybrid, in-office leaning. Careers: contextual.ai/careers. No open roles at time of writing.

Imbue is a San Francisco research lab building reasoning models and coding agents. It offers both in-person and fully-remote arrangements, which is unusual at this level. Careers: imbue.com/careers. All open roles currently marked specifically as being in SF.

World Labs, Fei-Fei Li’s spatial-intelligence startup, builds large world models that generate persistent 3D scenes. San Francisco, onsite. Careers: job-boards.greenhouse.io/worldlabs. Again of existing companies that I’m familiar with, World Labs ranks right at the top of my interests along with the Bio-adjacent orgs. Sadly all SF, although that’s not surprising.

A note for the curious: AI21 Labs (Tel Aviv, Jamba models) reportedly scaled back sharply in 2026 and its careers page showed no open roles when I checked, so confirm its status before investing time. Verify before applying: ai21.com/careers. I wasn’t overly thorough, but a few archive checks make me think it’s been empty like this for some time (maybe all of 2026, at least).

Together AI runs an “acceleration cloud” of GPU infrastructure plus inference and fine-tuning APIs. San Francisco, with roles in Amsterdam, London, Singapore, and India, mostly Bay Area-centric. Careers: job-boards.greenhouse.io/togetherai. No remote. Off to a shaky start lol.

Fireworks AI is a production inference cloud founded by Meta PyTorch and Google Vertex veterans. Redwood City, hybrid with some fully-remote US roles tagged on the board. Careers: job-boards.greenhouse.io/fireworksai. However, engineering seems expected to be onsite.

Baseten does mission-critical AI inference infrastructure. San Francisco, and the most openly remote-friendly of this group (work-from-anywhere, SF, or hybrid). Careers: jobs.ashbyhq.com/baseten. Brief browsing of the job descriptions makes it seem like they still largely expect engineers to be onsite for a large chunk of their time.

Modal is an AI-native serverless compute platform driven from Python. New York and San Francisco, hybrid. Careers: jobs.ashbyhq.com/modal.

Groq (custom LPU inference hardware, Mountain View, jobs.gem.com/groq), Cerebras (wafer-scale systems, Sunnyvale, job-boards.greenhouse.io/cerebrassystems), and SambaNova (custom accelerators, Palo Alto, sambanova.ai/company/careers) are the leading non-NVIDIA hardware plays, all hybrid and in-office leaning given the silicon work.

Databricks (Mosaic AI) trains, fine-tunes, and serves models on top of its data platform. San Francisco with offices worldwide, hybrid-first. Careers: databricks.com/company/careers. Snowflake (Cortex AI) serves models where customer data lives, is distributed, and runs a lighter in-office expectation. Careers: careers.snowflake.com.

Perplexity AI (answer engine, San Francisco, hybrid, jobs.ashbyhq.com/perplexity) and Cursor / Anysphere (AI code editor, San Francisco, mostly in-person, cursor.com/careers) are two of the highest-growth application companies, though both build on others’ models more than their own.

A couple of status notes worth knowing. Inflection AI pivoted to enterprise after Microsoft hired away its founders, and now operates as a B2B model company out of Palo Alto (inflection.ai/careers). Character.AI stopped building its own foundation models after Google’s licensing deal and now focuses on the product layer, with heavy Trust and Safety hiring (Redwood City, character.ai/careers). Both are useful examples of how fast these companies change shape.

The big-tech AI divisions are real options too, and tend to be more remote-tolerant than the pure frontier labs. Apple’s ML and AI group (Cupertino, hybrid, builds the Apple Foundation Models, jobs.apple.com) and Amazon AGI (covered above under the frontier labs) both train or heavily use frontier-scale models. NVIDIA gets its own section below.

Scale AI (training data and evals, San Francisco, scale.com/careers) and Hugging Face (the open-source model hub, fully remote and globally distributed, huggingface.co/careers) round out the infrastructure side. Hugging Face in particular is remote-first and central to the open ecosystem, which makes it a strong landing spot if relocation is off the table.

Allen Institute for AI (Ai2) is the standout champion of fully open AI: open weights and open data, code, and recipes (OLMo, Molmo, Tülu). Seattle, mostly in-office. Careers: allenai.org/careers.

EleutherAI is the nonprofit open-source lab behind The Pile and the Pythia suite, fully distributed, with a public Discord and a mentorship program, and an explicit “no PhD, no problem” stance. Get involved: eleuther.ai/community.

Nous Research is a leader in American open-source models (the Hermes series) and distributed training (Psyche), fully remote. It even runs a real jobs page, which is rare for a collective: nousresearch.com/careers.

LAION is the German nonprofit behind the LAION-5B dataset that trained Stable Diffusion, a distributed volunteer network rather than an employer. Get involved: laion.ai.

Prime Intellect builds decentralized training infrastructure that lets anyone contribute compute to open model runs. San Francisco, hybrid with some remote-eligible roles. Careers: primeintellect.ai/careers. Stability AI survived its 2024 turmoil and continues on generative creative tools out of London, hiring selectively (stability.ai/careers).

• Stanford: HAI, SAIL, and CRFM, the group that coined “foundation models.”
• UC Berkeley BAIR, whose Sky Computing Lab is where vLLM originated.
• MIT CSAIL, CMU MLD and LTI, and Princeton PLI, which runs a sizable academic GPU cluster and is the most remote-tolerant of the set (its research-software-engineer posting notes flexible arrangements).
• ETH Zurich and EPFL’s Swiss AI Initiative, which shipped the fully open multilingual Apertus model: ai.ethz.ch, ai.epfl.ch.
• Mila in Montreal, founded by Yoshua Bengio, with the clearest formal jobs page of the academic set: mila.quebec/en/recruitment.

On the US national labs, Argonne (the Aurora supercomputer and AuroraGPT), Oak Ridge (Frontier), and Lawrence Berkeley (NERSC) all do frontier-scale AI for science, mostly on-site, with a shared intern pipeline through DOE SULI.

NVIDIA is the company the rest of this post quietly depends on. Its GPUs and the CUDA stack run almost every frontier training run, and beyond the silicon it ships the dominant inference software (TensorRT, the Dynamo serving stack), maintains a large research org in NVIDIA Research, and trains its own open models (the Nemotron family). Santa Clara, with research spread across many sites. It is a notable return-to-office holdout, with location driven by the team rather than a company-wide mandate, which makes it more remote-tolerant than most of the labs. Careers: nvidia.com careers, research roles at research scientists.

AMD is the main GPU challenger (the Instinct MI accelerators and the ROCm software stack). The interesting angle for an outsider is that ROCm is the open competitor to CUDA and is still catching up, which means a large, public surface area you can contribute to, in exactly the low-level systems work frontier labs hire for. Santa Clara. Careers: careers.amd.com.

The GPU clouds (neoclouds) rent frontier-scale capacity and hire heavily for infrastructure, networking, and serving engineers, often with more remote flexibility than the labs. The notable ones: CoreWeave (coreweave.com/careers), Lambda (lambda.ai/careers), and Crusoe (crusoe.ai/about/careers). The hyperscaler clouds (AWS, Google Cloud, Azure) run their own large AI-infrastructure orgs too, though those blur into the big-tech divisions above.

• Roofline: An Insightful Visual Performance Model (Williams, Waterman, Patterson, 2009). Arithmetic intensity, and whether you’re compute-bound or memory-bound.
• The CUDA C++ Programming Guide (NVIDIA). The execution and memory model every later paper assumes.
• Online normalizer calculation for softmax (Milakov, Gimelshein, 2018). The streaming-softmax trick that makes Flash Attention possible. Read it first.
• Triton (Tillet, Kung, Cox, 2019). Tile-based GPU kernels in Python, the practical entry point.
• FlashAttention, FlashAttention-2, and FlashAttention-3 (Dao et al.). The landmark IO-aware kernel and its evolution toward async pipelining and low precision.
• ThunderKittens (Spector et al., 2024). A tile-based embedded DSL, the modern way to write competitive kernels.

• LLM.int8() (Dettmers et al., 2022). The outlier story: why naive INT8 breaks and how mixed-precision fixes it.
• SmoothQuant (Xiao et al., 2023). Migrate quantization difficulty from activations into weights.
• GPTQ (Frantar et al., 2022) and AWQ (Lin et al., 2024). The two workhorse post-training methods, worth contrasting.
• QuIP, QuIP#, and QTIP (Cornell). The incoherence-processing line that pushed weight-only quantization toward two bits.
• AQLM (Egiazarian et al., 2024). The additive-codebook branch to contrast with QuIP#.
• SnapKV (2024). KV-cache compression, essential for long-context serving.
• For low-precision number formats: FP8 Formats for Deep Learning, FP8-LM, and the recent push into FP4 training and NVFP4 pretraining.

• Scaling Laws for Neural Language Models (Kaplan et al., 2020). The bedrock.
• Training Compute-Optimal LLMs (Chinchilla) (Hoffmann et al., 2022). The correction: scale model and data together.
• muP / Tensor Programs V (Yang et al., 2022). Tune hyperparameters on a tiny proxy and transfer them zero-shot.
• GSPMD (Xu et al., 2021) and PaLM (Chowdhery et al., 2022). Compiler-based sharding, and the canonical large-scale case study (and the standard MFU definition).
• The JAX Scaling Book (Austin et al., 2025). The best practical primer on rooflines and parallelism. This is the one Vlad tells you to work through exercise by exercise.
• The DeepSeek-V3 Technical Report (2024). The modern recipe end to end.

• Large Scale Distributed Deep Networks (Dean et al., 2012). The parameter server and async SGD, the origin of the gradient-staleness story.
• GPipe and PipeDream. Synchronous and asynchronous pipeline parallelism, worth reading as a pair.
• Megatron-LM (Shoeybi et al., 2019). The standard for tensor parallelism.
• ZeRO (Rajbhandari et al., 2019) and PyTorch FSDP (Zhao et al., 2023). Sharding optimizer states and parameters, in theory and in production.
• Reducing Activation Recomputation (Korthikanti et al., 2022) for sequence parallelism, and Ring Attention (Liu et al., 2023) for context parallelism.

• Orca (Yu et al., 2022). Continuous batching and iteration-level scheduling.
• PagedAttention / vLLM (Kwon et al., 2023). OS-style paging for the KV cache, the foundation of vLLM.
• Speculative decoding (Leviathan et al., 2022) and the DeepMind derivation (Chen et al., 2023), plus Medusa (2024) for self-speculation.
• Splitwise (2023) and DistServe (2024). Disaggregating prefill and decode, the reference for modern serving.
• For the efficiency metrics, MFU is defined in PaLM and MBU in this Databricks engineering post.

• Reinforcement Learning: An Introduction (Sutton and Barto), and OpenAI’s Spinning Up as the code-first companion.
• Deep RL from Human Preferences (Christiano et al., 2017) and PPO (Schulman et al., 2017). The roots of RLHF.
• InstructGPT (Ouyang et al., 2022). The full SFT to reward-model to PPO pipeline.
• DPO (Rafailov et al., 2023). Preference optimization without the RL loop.
• DeepSeekMath (introduces GRPO) and DeepSeek-R1. The RL-for-reasoning work everyone is building on now.

• Outrageously Large Neural Networks (Shazeer et al., 2017). The foundational sparse MoE layer.
• GShard (2020) and Switch Transformers (2021). Expert parallelism and simplified routing at scale.
• Tutel (2022) for the systems side, and Mixtral (2024) and DeepSeekMoE (2024) for concrete modern designs.

• Halide (Ragan-Kelley et al., 2013). Decoupling algorithm from schedule, the idea under most ML compilers.
• TVM (Chen et al., 2018) and XLA / OpenXLA. The full compiler stacks behind PyTorch and JAX.
• Triton and ThunderKittens again, as the DSLs people actually write kernels in.
• Mojo (Modular, from MLIR and LLVM creator Chris Lattner). A Python-superset systems language that compiles through MLIR toward hand-tuned-kernel performance, betting that one language can span Python ergonomics and GPU-kernel speed.
• Parrot (NVlabs). A header-only C++20/CUDA library that uses expression templates for implicit fusion and lazy evaluation, composing element-wise math, reductions, and reshapes without materializing intermediates. A library-level take on the same avoid-the-slow-memory idea behind Flash Attention. I’ve actually landed a few PRs in the Parrot library recently.

• Distilling the Knowledge in a Neural Network (Hinton et al., 2015). Soft targets and temperature, the origin.
• Sequence-Level Knowledge Distillation (Kim, Rush, 2016) and DistilBERT (Sanh et al., 2019).
• MiniLLM (Gu et al., 2023) on reverse-KL distillation, and Thinking Machines’ more recent On-Policy Distillation for efficient reasoning transfer.

• ReAct (2022), Toolformer (2023), and Reflexion (2023). The reasoning-and-acting loop, learned tool use, and self-correction.
• Voyager (2023) for open-ended skill acquisition, and SWE-bench (2023) for how coding agents are actually measured.
• FunSearch (2023) and AlphaEvolve (2025). Search over programs, the line Vlad points at for the work above the model.

• vLLM (Python, CUDA, C++). The reference high-throughput serving engine, and the single highest-signal repo on this list. Enormous surface area for model support, kernels, and scheduling.
• SGLang (Python, CUDA, Rust). LMSYS’s production serving engine, co-equal with vLLM in hiring signal and very active.
• TensorRT-LLM (Python, C++/CUDA). NVIDIA’s optimized stack; contributing signals you can work at the kernel-runtime boundary.
• llama.cpp (C/C++). The de-facto local inference engine. Note the org moved from ggerganov to ggml-org.
• LMDeploy, MLC-LLM (compiler-based deployment), and ExLlamaV3 (use V3, not the now low-maintenance V2).
• Skip TGI as a contribution target; it was archived in 2026 and the ecosystem moved to vLLM and SGLang.

• Triton (Python DSL). The dominant kernel DSL. Note it moved from openai/triton.
• CUTLASS (CUDA C++). Deep, hard, and respected; real tensor-core expertise.
• FlashAttention (CUDA / CuTeDSL). About as load-bearing as a kernel repo gets.
• ThunderKittens (CUDA tiles) and Liger-Kernel (Triton). Lower barrier to writing fast kernels, and Liger is an especially friendly first stop for Triton training kernels.

• llm-compressor (Python). The vLLM-ecosystem PTQ library, the most strategically current quant repo.
• GPTQModel, the active successor to the now-deprecated AutoGPTQ.
• torchao (PyTorch-native quant and sparsity) and bitsandbytes (the ubiquitous k-bit library).
• HQQ for calibration-free quantization. Skip AutoAWQ (archived) and AutoGPTQ (deprecated).

• Megatron-LM (Python, CUDA). The reference large-scale training framework.
• DeepSpeed (ZeRO and offload; org moved to deepspeedai) and torchtitan, the best entry point for modern FSDP and parallelism.
• Axolotl and Unsloth (the latter bridges training and Triton kernels). LitGPT and EleutherAI’s GPT-NeoX for clean from-scratch implementations.
• nanochat, Karpathy’s full ChatGPT-pipeline-from-scratch. Use this rather than the now-deprecated nanoGPT, as a learning and portfolio project.

• JAX (note the org is jax-ml), Flax, Optax, and Equinox.
• MaxText, Google’s scalable pure-JAX LLM reference, and Levanter for bitwise-reproducible training. Both signal TPU-scale training skill.

• TRL (the default open post-training library, GRPO/DPO/PPO), verl (one of the most-used scalable RL stacks), and OpenRLHF (Ray and vLLM based).
• alignment-handbook for recipe and reproduction contributions. Skip trlx, which is effectively stale.

• OpenHands (leading open SWE-agent; org moved from All-Hands-AI), DSPy (programming, not prompting), and smolagents (a tiny, readable codebase ideal for a first contribution).
• LangChain and LlamaIndex for applied agent work.
• For evals: lm-evaluation-harness (the de-facto backend), lighteval, and UK AISI’s Inspect. Inspect’s evals repo explicitly takes community-contributed evals, a clean way to land a frontier-adjacent merged PR.

• tokenizers (Rust plus Python bindings), a genuinely good place to show Rust and performance skill.
• datatrove (the engine behind FineWeb), nanotron (minimal 3D-parallelism pretraining, with a real contributor guide), and Ai2’s dolma.