I'm an ML engineer/ML research engineer looking for roles where there are gradients to descend, models to build, and/or datasets to gather. In general I want to advance the frontier of what's possible. My credentials are more towards research engineering than applied work, but I'm open to both. My last job was as a contract research engineer doing AI safety work at Redwood Research. I spent about half my time on our synthetic datasets of safe & unsafe LLM behavior, and the other half on fancy finetuning. I finetuned an LLM to exhibit unsafe behavior while fooling logistic probes trained to detect it, but only in deployment. Weird multi-part composite loss. My other big project is independent research in text-to-image models that are trained solely on unlabeled image data, exploiting CLIP embeddings for the link between text and images. Currently fighting with spherical flow matching models and racing the ICLR paper deadline.
If you need help getting ML systems up and running, getting them fast and correct, or if any of that stuff sounds interesting to you, get in touch.
I'm an ML engineer/ML research engineer looking for roles where there are gradients to descend, models to build, and/or datasets to gather. In general I want to advance the frontier of what's possible. My credentials are more towards research engineering than applied work, but I'm open to both. My last job was as a contract research engineer doing AI safety work at Redwood Research. I spent about half my time on our synthetic datasets of safe & unsafe LLM behavior, and the other half on fancy finetuning. I finetuned an LLM to exhibit unsafe behavior while fooling logistic probes trained to detect it, but only in deployment. Weird multi-part composite loss. My other big project is independent research in text-to-image models that are trained solely on unlabeled image data, exploiting CLIP embeddings for the link between text and images. If nothing goes horribly wrong I'll be submitting a paper for ICLR in about 7 weeks.
If you need help getting ML systems up and running, getting them fast and correct, or if any of that stuff sounds interesting to you, get in touch.
I'm an ML engineer/research engineer looking for work in that vein. I can train models, I can design architectures, I can implement papers, I can gather real and synthetic data, and I can in general solve the problems that come up when trying to build ML systems. Background: I recently ended a 3 month contract at Redwood Research doing AI safety research where I spent about half my time improving our synthetic data pipelines and about half finetuning LLMs to exhibit unsafe behavior in deployment but not training, and to do so while defeating probes trained to detect the unsafe behavior (but only in deployment). I also have an ongoing solo research project working on text-to-image models trained without any text, relying on embeddings for conditioning. [1] is a (very out of date) blog post about that. There'll be a paper soon inshallah. Before the ML stuff I worked at a startup where we built a new cryptocurrency, and before that another startup.
Location: New York, NY
Remote: Either
Willing to relocate: No
Technologies: Machine learning, JAX, PyTorch, Python, Rust, Haskell, OCaml
Resume: https://www.echonolan.net/resume/cv.pdf
Email: echo@echonolan.net
I'm looking for an ML engineering gig. I can help you gather data, preprocess it, design models, train models, etc etc. My ideal job would be doing generative AI stuff with images/audio/video, but I'm open to anywhere there's gradients to descend. Recently I've been working on a project building a text-to-image model that learns solely with unlabeled image data, relying on CLIP for the link between captions and images[1]. I think it's a) cool and b) demonstrative of strong abilities. At a higher level of abstraction you can think of this as embedding guided content synthesis. The model learns to generate images conditioned on their CLIP embedding being within an input spherical cap. If you center the cap on the CLIP embedding of some text you get images that look like they'd have that caption, if you center it on the embedding of another image you get semantically similar images. The radius of the cap determines how similar the outputs are.
There's a listing for "Junior Software Engineer", but I don't see anything for "Software Engineer" or "Ex-founder" on the page you linked. And the salary ranges seem implausibly low, at least for Americans.
The low-end salary on most of those positions is below the US federal minimum wage. I realize the company is hiring "global" candidates, but if what you really mean is "candidates from outside the US, Canada, or Europe", that's unusual enough that you should be calling it out.
It's even worse. Let's assume good faith and assume the below is a typo that will be immediately rectified:
> Dallas, TX, US $10K - $30K
In the USA, this is below the poverty line, and less than what you make wiping down floors in a fast food burger place off the highway in a shady part of town with no health insurance.
This is a part-time position, expected to be concurrent with full-time employment or higher education. It's paid hourly and the salary range was defined based on expected number of hours, estimated about 1/10 to 1/5 of a full-time position (there's no way to define hourly compensation in that form field, and the number of hours isn't known exactly ahead of time).
You should just state the hours (4-8 hours a week apparently) and hourly rate (still no idea based on this vague math) in the job req then. Especially when listed next to other offshore jobs with such low wages, it's a confusing and bad look.
If multiple unrelated people are so mislead that you have to answer questions about it, it's a good sign the issue is your communication, not their understanding of it.
My optimistic hope when reading their salaries is that they posted a monthly value instead of annual. If the values are in fact annual then the execs are completely delusional.
Can't be. Nobody would ever pay n entry level data center technician in Dallas $30k per month either. It's bizarre. You'd have to hire exclusively felons without high school degrees to be able to command such a low wage.
We get some good candidates when the low end isn't too high because some people are just starting their career, are looking for internships, or are humble. We've hired some very talented people that come in with low expectations, and then we're able to calibrate them upwards.
Salaries vary widely around the world, as you know, and we hire in almost all countries. We have team members in 22+ countries, including US and EU. If you tried hiring globally I expect you'll find many talented people won't apply for your positions if the salary range is too far out of their expectations.
Location: NYC
Remote: Either on-site or remote is fine
Willing to relocate: No
Technologies: Machine learning, generative AI, text-to-image models, JAX, Python, Rust, PyTorch, OCaml, Haskell
Résumé/CV: https://www.echonolan.net/resume/cv.html (HTML) or https://www.echonolan.net/resume/cv.pdf (PDF)
Email: echo@echonolan.net
Ideally, I'd find a job doing ML engineering on text-to-image, text-to-video, text-to-audio or related models - recently I've built a text-to-image model that is trained with unlabeled images alone, using CLIP for the link between captions and images[1]. I'm interested in ML engineering in other domains as well, and my last job was building a new cryptocurrency, so I have skills there too.
Here's my blurb about the model I built:
Recently, I've built a text-to-image model that is trained without any text labels, using unlabeled images and CLIP for the link between captions and images. This has never been done or even investigated before. Results are promising so far. I think this work is the best representation of what I'm capable of. In the process, I gathered a dataset of 33 million images for training data, including removing redundant images, deduplicating, and taking stills from any videos. I ported a VQGAN implementation from PyTorch to JAX, built an efficient preprocessing pipeline, built transformer models in JAX, and designed and trained baseline models and more sophisticated ones. To support the approach I eventually settled on, I designed and implemented an efficient algorithm to sample unit vectors from a finite set, conditioned on the vectors being inside a spherical cap. For that I needed to write a Python library in Rust to help with constructing the space partitioning data structure used for sampling. The sampling algorithm gets used to generate training examples and the model learns to sample images conditioned on the image's CLIP embedding being within an arbitrary spherical cap.
> You think the outfit that took down Weinstein cares about some SV tantrums?
Metz cares about access to SV sources, probably. I'll quote Nostalgebraist[0]:
> The journalist also written a soon-to-be-published book about AI work at “Google, Microsoft, Facebook and OpenAI,” whose blurb makes it sound impressed with its subjects, and also touts his “exclusive access to each of these companies.” So, this is someone whose career depends on being in the good graces of the big-name Silicon Valley crowd, and presumably cares a lot whether e.g. Paul Graham is mad at him.
I don't think you can do the sort of work Cade Metz wants to do if the first thing anyone from OpenAI thinks of when you write to them is "oh, it's that asshole that doxed Scott Alexander".
Note the tweet says they pay 30% more than "traditional retail stores". I don't think that's the right comparison. Pick & pack in a warehouse is much more physically demanding than being on the floor at Best Buy. The better question is do they pay more than UPS, or for that matter, picking strawberries?
Remote: Yes, but prefer in-office
Willing to relocate: Yes, to the Bay Area
Technologies: machine learning, transformers, LLMs, finetuning, flow matching, JAX, PyTorch, Python
Resume: https://www.echonolan.net/resume/cv.pdf
Email: echo@echonolan.net
I'm an ML engineer/ML research engineer looking for roles where there are gradients to descend, models to build, and/or datasets to gather. In general I want to advance the frontier of what's possible. My credentials are more towards research engineering than applied work, but I'm open to both. My last job was as a contract research engineer doing AI safety work at Redwood Research. I spent about half my time on our synthetic datasets of safe & unsafe LLM behavior, and the other half on fancy finetuning. I finetuned an LLM to exhibit unsafe behavior while fooling logistic probes trained to detect it, but only in deployment. Weird multi-part composite loss. My other big project is independent research in text-to-image models that are trained solely on unlabeled image data, exploiting CLIP embeddings for the link between text and images. Currently fighting with spherical flow matching models and racing the ICLR paper deadline.
If you need help getting ML systems up and running, getting them fast and correct, or if any of that stuff sounds interesting to you, get in touch.