About the role
The pretraining team trains the base models that every other Clexerl system builds on. The work is empirical and comparative: architectures, optimisers, data recipes and training schedules are tested against each other at small scale, then promoted to larger runs only when the evidence holds. The question you chase is simple to state and hard to answer: what makes the next model better per unit of compute?
Frontier AI feeds Clexerl's own products, so a better base model shows up in Clexerl Go's agents, in long-term memory and in multimodal understanding within a release cycle. That gives the team direct signal on what matters downstream, and it keeps scaling-law work honest, because a loss curve that does not translate into a capability is not a result.
You work in small groups with a research engineer alongside every scientist. Experiments are written up in full, including the ones that failed, and claims are reviewed by someone who did not run them. Disagreement happens in the write-up, then the group commits to a plan.
What you’ll do
- Design and run ablations on architecture, optimiser, initialisation and learning-rate schedule, and fit scaling laws that predict behaviour at the next order of magnitude.
- Propose hypotheses about training dynamics, such as loss spikes, representation collapse or data-ordering effects, and test them with controlled experiments.
- Work with the data team to evaluate mixtures and curricula, and measure how each source changes downstream capability rather than only validation loss.
- Write research code in PyTorch or JAX that scales cleanly from a single accelerator to a multi-node run.
- Review other people's experiments and analysis, and flag confounds such as seed variance, contamination or mismatched token budgets.
- Present results in writing to the wider Frontier AI group and recommend which ideas deserve a larger compute allocation.
What you’ll bring
- A strong research record in deep learning, shown through papers, open-source work or large-scale experiments you can explain in detail.
- Fluency with transformer internals: attention variants, normalisation placement, positional encodings, mixture-of-experts routing and their effect on stability.
- Practical command of optimisation, including Adam-family behaviour, weight decay, warmup, and how hyperparameters transfer across width and depth.
- Good statistical habits: you know when a difference is inside the noise and you size experiments accordingly.
- Comfort writing and debugging training code in PyTorch or JAX, including distributed data loading and checkpointing.
- Clear written communication that lets a colleague reproduce your experiment without asking you questions.
Nice to have
- Experience with muP-style parameterisation or other hyperparameter transfer methods.
- Prior work on sparse or mixture-of-experts models at scale.
- Published work on scaling laws, data efficiency or training dynamics.
- Familiarity with low-precision training in FP8 or similar formats.
Work you might take on
- 01
Fit scaling laws for a new mixture-of-experts configuration and predict its loss at ten times the compute before committing a run.
- 02
Investigate recurring loss spikes in a mid-scale run and isolate whether the cause is the optimiser state, the data order or numerics.
- 03
Compare three learning-rate schedules, including warmup-stable-decay, and publish which one transfers best across model widths.
- 04
Build an ablation suite that measures how each architectural change affects long-context retrieval, not just perplexity.
Compensation
$340K – $560K
The US base salary range for this role. Every offer also includes equity in Clexerl and, when it spins out, in the venture you help build. We tell you the reasoning behind your number.
Benefits
Health, fully covered
Medical, dental and vision for you and your dependants, premiums paid in full.
Two kinds of equity
Equity in Clexerl, plus a stake in the venture you help build when it spins out.
22 weeks parental leave
Fully paid, for every parent, however your family arrives.
Rest that sticks
Flexible time off with a four-week minimum we actually check on, plus a company-wide week in December.
Learning and compute
A yearly budget for courses, conferences and books, and compute credits for side projects.
401(k) with a 4% match
Matched from your first paycheque and vested immediately.
How we hire
Apply 5 minutes
Send links to what you have built and a few lines on why this role. No cover letter needed.
Intro call 30 minutes
A conversation with the hiring manager about the work, the team and what you want next.
Work sample 2–4 hours
A paid take-home or a pairing session on a real-shaped problem. You pick the format.
Studio day One day
Meet the team, go deep on your work and ours. Remote or in person.
Offer Within 5 days
A written offer with the full range, the equity and the reasoning behind both.
Opens an email with a short template. Links beat résumés.