hlfshell

  Location: San Diego, CA, USA or Remote
  Remote: Preferred but not required
  Willing to relocate: Yes
  Technologies: Python, golang, AI, robotics, ROS, GCP, AWS, Terraform, Pulumi, React, Vue, Typescript, Computer Vision, Reinforcement Learning, tensorflow, pytorch
  Resume: https://github.com/hlfshell/resume
  Portfolio: https://hlfshell.ai/projects
  Current startup: https://hiredcoach.ai
  Email: keith@hlfshell.ai

I am an experienced systems-focused software engineer who builds and integrates cloud solutions, AI tooling, composable agent frameworks, and robotics-oriented software, capable of building either the low level pieces or of full on orchestration for AI workflows. I have a strong background in cloud SaaS products, robotics, reinforcement learning, AI, and more.

ARC Raiders runs fine with anticheat on Linux. As does the Finals.

Just launched a startup/life style business where I use AI to help people practice for upcoming interviews - https://hiredcoach.ai

Already have been told by some users that the interview prep they got from it has correctly predicted several of the actual interview questions they got, crediting its prep for their breezing through the interview rounds.

I'm really hoping it helps a lot of people!

What's missing from this explanation is that the corporate tax rate was also much higher, but R&D dramatically cut down profit that would be taxed and was taxed lower. So large corporations like Bell Labs and co would basically say "do we give the government X in taxes, or do we spend X on research?". They chose research, so we got the technology that powers our world.

That, combined with stock buybacks and the general take over of Friedman-economics resulted in a far more focused short term thinking and outsourcing research as much as possible due to uncertain horizon risks.

Pretty addicting game!

No. We can add guardrails, but nothing that has been proven to work reliably.

Also being utilized in modern VLA/VLM robotics research - often called "Constitutional AI" if you want to look into it.

The generally accepted term for the research around this in robotics is Constitutional AI (https://arxiv.org/abs/2212.08073) and has been cited/experimented with in several robotics VLAs.

There's shilling a product because someone handed you a bag, and then there's building a product you believe in. You feel okay with it because it's clearly the latter versus another NordVPN commercial. Even if the product ends up failing (and I am under no predilection to believe this will) he has presented nothing but honest enthusiasm towards his goal that you can't help but root for it.

He's gone all in on the Crunch Labs brand, which is kind of built around the younger audience. This isn't a bad thing, but it does mean that older edutainment enjoyers kind of age out of his stuff. Not to say there's no value in them, but there will be more of an entertainment focus than prior edutainment focused videos.

I recommend checking out Stuff Made Here; great build videos of engineering principles in an entertaining fashion to show building cool complicated stuff.

Xyla Foxlin, a wonderful maker, also posts educational videos between her projects, like an in-depth look at how plane wings work.

Taking a break from my agentic AI framework for prototypes and makers arkaine(1) and made two fun useful apps for myself

1. Eli5 equations(2) uses an LLM to convert a given picture of an equation to latex and, if given additional context, breaks down the equation parts to explain it. Gemini for the model.

2. reflecta - a journal prompting app with deepseek to help reword and target the prompts towards you better.

(1) https://arkaine.dev

(2) https://eli5equation.com

(3) https://reflecta.hlfshell.ai

Wealth concentration and corporate consolidation.

It's only a problem if we live in a society where people's perceived value, and thus capability of living a healthy, full life, is tied to their productivity to produce profit for an increasingly shrinking pool of people and organizations.

Which is what we have, hence the problem.

Yes, AI has the potential to screw things up royally. But do not mistake its' exacerbation of symptoms as the true illness.

The issue is that as music progresses and changes so too does distribution networks. Traditional, or even nontraditional to those from the pre spotify internet days, pipelines of music discovery have been largely co-opted by industry. Outlets of organic discovery are different now - and people typically don't continually keep changing their habits enough to keep up with it.

Pair this with the fact that most people settle their musical tastes to be in line with when they are experiencing the most emotionally significant time in their early lives (high school for some, college for others, etc) and the result is an assumption that

A) What they encounter forms an overall opinion of "all" new music despite being the tip of the iceberg and

B) It's not as good as what they grew up on

Google (Alphabet's) stock price has generally gone up 200% in the past 5 years. That is the only reason he is there, and that is the only way he is judged.

Google is very much suffering from the classic Innovator's Dilemma [1]; a side effect of being too focused on stock price and not long term planning.

A better management with long term thinking would utilize Google's enormous base of talented engineers far better.

[1]https://www.hbs.edu/faculty/Pages/item.aspx?num=46

Reminds me of Bonfire from... Jeez 2009. https://youtu.be/O3MZYRAZJNk?si=C9fcjABHWwOxbs5O

I've seen this being researched under the term Constitutional AI, including some robotics papers (either SayCan or RT 2? Maybe Code as Policies?) that had such rules (never pick up a knife as it could harm people, for instance) in their prompting.

The reason the time (blitz) games make sense is because the distilled functionality is of a 50ms Stockfish eval function. The engine likely would perform worse as only the human would benefit from the additional time.

As for limited search tree I like the idea! I think it's tough to measure, since the time it takes to perform search across various depths vary wildly based on the complexity of the position. I feel like you would have to compile a dataset of specific positions identified to require significant depth of search to find a "good" move.

I did a talk about this! (And also wrote up about my talk here[1]). This paper is a great example of both knowledge distillation. It's less of a paper about chess and more about how complicated non linear search functions - complete with whatever tuning experts can prepare - can be distilled into a (quasi-linear, if it's a standardized input like chess) transformer model.

[1]: https://hlfshell.ai/posts/deepmind-grandmaster-chess-without...

It kind of exists; here's one such video demonstrating the (not as advertised, but still kinda works) results on a microphone.\

https://www.youtube.com/watch?v=FyeCn7HlLck

Younger ears find it excruciating though.

Location: San Diego, CA

Remote: Preferred unless it's a great fit

Willing to relocate: Yes, for the right offer

Technologies: PyTorch, Tensorflow, Python, Golang, AWS, GCP, Kubernetes, Robotics, LLMs

Resume/CV: https://github.com/hlfshell https://github.com/hlfshell/resume

Site/Portfolio: https://hlfshell.ai

Email: kchester@gmail.com

12 years of experience doing backend systems, including five at robotics startups doing data pipelines and automated deep learning training pipelines. Completing my Masters of Science in Robotics Engineering; actively looking for a new role in robotics, AI, or other interesting spaces. Currently investigating language enriched models and AI w/ reinforcement learning!

In my experience using LLMs like CoPilot:

Web server work in Go, Python, and front end work in JavaScript - it's pretty good. Only when I try to do something truly application specific that it starts to get tripped up.

Multi threading python work - not bad, but occasionally makes some mistakes in understanding scope or appropriate safe memory access, which can be show stopping.

Deep learning, computer vision work - it gets some common pytorch patterns down pat, and basic computer vision work that you'd typically find tutorials for but struggles on any unique task.

Reinforcement learning for simulated robotics environments - it really struggles to keep up.

ROS2 - fantastic for most of the framework code for robotics projects, really great and recommended for someone getting used to ROS.

C++ work - REALLY struggles with anything beyond basic stuff. I was working with threading the other day and turned it off as all of its suggestions would never compile let alone do anything sensible.

I was speaking about LLM weights specifically and llama, not all models and work at FAIR.

Per your links, it's clear that FAIR does have a good history of open source work.

It started by accident, with the original llama weights being leaked by two separate employees. They've since embraced opening the weights, which I'm all for.

As for why? I have a theory: Meta is not in a position to capitalize upon the model itself. Yes, they can use it internally, and maybe their competitors can copy it to - but there are no real competitors to Facebook or Instagram that can benefit from it enough to make it a differentiating facet.

Thus, releasing stuff for open source does two things:

1) Make them more attractive to research talent (Apple famously recently started to publish research because their traditional secrecy was causing issues with hiring top talent) and...

2) Continues to undermine the ability to make $$$ off of model alone, driving it towards being a commodity rather than the long term profit engine for other companies.

Rule of thumb - if it runs well on steam deck / steamos it probably doesn't have anything too extreme... most multiplayer games won't run on the windows compatibility layer as it allows the client a possible level of control that would look similar to a low level cheat.

Nearly every competitive game game is going to be utilizing some form of anticheat like this, so single player games tend to not have this issue (drm is a separate conversation)

This is why smaller AA games still keep this alive - to build community. For instance: Squad, a 80-100 player multiplayer game which primarily depends upon privately hosted servers ran by communities that actually care about the community and game. You'll still run into jerks like any online game, but it's a world of difference to the one-match/single-serving strangers you usually play with in other games.

Location: San Diego, CA

Remote: Preferred unless it's a great fit

Willing to relocate: Yes, for the right offer

Technologies: PyTorch, Tensorflow, Python, Golang, AWS, GCP, Kubernetes, Robotics, LLMs

Resume/CV: https://github.com/hlfshell https://github.com/hlfshell/resume

Site/Portfolio: https://hlfshell.ai

Email: kchester@gmail.com

12 years of experience doing backend systems, including five at various robotics startups doing data pipelines and automated deep learning training pipelines. Completing my Masters of Science in Robotics Engineering; actively looking for a new role in robotics, AI, or other interesting spaces. Currently investigating language enriched models and AI w/ reinforcement learning!

I've worked with this poster before for a few years before moving on to break into a different industry. Fantastic engineer, great person to work for. I highly recommend sending your resume their way.