yorick

Datakami | ML Engineer | Netherlands (within 2hrs of Nijmegen) | Remote (NL, DE, BE) | No visa sponsorship | https://datakami.com

We're a software engineering consultancy specialized in generative AI. Small team of senior ML engineers offering remote expertise to startups across US/Europe, solving unique challenges of running generative AI in production.

Looking for ML engineers who are a mix of ML expert, software engineer, researcher, and hacker. You'll work embedded in client projects on: - Converting bleeding-edge open source AI models to production - Building production LLM pipelines from scratch - Improving models on speed, robustness, and performance - Designing custom LLM benchmarks and evaluation - Building and scaling ML infrastructure - Setting up monitoring, tracing, and prompt management

Tech stack: Python, LLMs, AWS/GCP, MLOps tools, Docker, Git, Nix

We value: self-starters, quick learners, strong communication skills, software quality, open source. Role involves engineering, talking to clients and outreach.

Requirements: Strong Python, experience with production LLM systems, cloud platforms, MLOps. Must be EU work eligible and live within 2 hours of Nijmegen, Netherlands. Remote work with occasional meetups.

Benefits: 25 days PTO, home office budget, professional development budget

Apply: https://datakami.com/careers

Recruiters/freelancers/agencies: we're not working with recruiters or considering freelancers or agencies at this time.

It looks like you can use the gemma tokenizer to count tokens up to at least the 1.5 models. The docs claim that there's a local compute_tokens function in google-genai, but it looks like it just does an API call.

Example for 1.5:

https://github.com/googleapis/python-aiplatform/blob/main/ve...

To provide some counterweight to all the overwhelmingly positive reviews:

I've used kagi for 6 months and have over 7500 searches with them. It mostly works, but there are a few downsides compared to Google:

- The latency is a lot higher than google, taking over a second to display any results. - The results are often not as relevant, I have to frequently retry my search in Google. - The results for anything local (I'm not in the US) are abysmal. Searching for anything in my city instead only gives me results for the city's history.

Still, I persist in using Kagi, mainly because it's not Google and I want them to succeed. The results are frequently good enough for me to stay with them.

Other comment: https://news.ycombinator.com/item?id=41813058

To passively radiate a gigawatt of heat in space at 100C, you'd need a radiator with a (visible) surface area of 1 million square meters.

It uses the 60ghz radar from https://vayyar.com/ . I'm not sure how or if it works, I haven't tried it myself. They also sell dev kits: https://walabot.com/products/walabot-developer-pack-new

2.4ghz isn't great at detecting small obstacles like wires. There's a smartphone mounted device that can do this with 60ghz: https://walabot.com/

This should be possible, it already exists for Google Photos: https://www.stg-uploader.xyz/

More about this on wikipedia: https://en.wikipedia.org/wiki/Variability_hypothesis

IMHO the evidence points to this being a cultural effect that can be changed with policy.

Adding `busybox` and `bashInteractive` to the container contents gives you enough of a comfortable environment to work in without losing too much space.

note: ipv6 is 128 bits, which should, in fact, be enough for everybody

I'm not sure that the answers that the model provides have anything to do with what it's actually doing. The way they seem to be prompting it also exhibits this issue, where they first have it arrive at a conclusion and then come up with an explanation for this conclusion. LLMs do not have an inner voice to reason with, and tokens generated later do not influence earlier tokens (unless you're doing beam search, but you mostly aren't). It would be much improved if asked to do reasoning first and then arrive at a conclusion.

Same issue on sway + nvidia, it can be worked around by disabling wayland support for electron apps. Not ideal.

https://artemis.sh/2022/09/18/wayland-from-an-x-apologist.ht...

talks about it at length.

It seems the main things missing for Talon are:

- input emulation, doable via uinput but not great - standard way to query the list of windows and active focus - for dwell-click support, you need to be able to know if the user is moving their mouse or clicking so you can cancel your autoclick

Sourcegraph seems to have collected a bunch of these (now leaked) email addresses from signups on self-hosted instances.

I remember being very surprised when I was signed up to their mailing list after I made an account on my self-hosted instance, and I'm not sure about the ethics (and legality) of collecting these in the first place.

SEEKING WORK | Europe (CET) | REMOTE

We're a 2 person team doing on-demand research, development and prototyping, specializing in generative AI & LLMs.

Our services include: AI Strategy, custom LLMs, prompt engineering, NLP, dataset vetting

Recently used technologies: langchain, llama-index, python, datasette, llama.cpp, stable diffusion, typescript

Website: https://datakami.nl/

Email: hello@datakami.nl

I've had poor experiences with Wasabi's availability in practice. Does cloudflare deal with hours-long outages well? Do you know if B2 fares better?

These ways already exist: [waypipe](https://gitlab.freedesktop.org/mstoeckl/waypipe/) is fairly complete already.

You may want to consider running some proof-of-work in the browser, that would be costly to run for bots.

Telegram does not require your phone number, you can set it up with just a user name.

A patch is in the works for ARM chips as well (http://lists.infradead.org/pipermail/linux-arm-kernel/2017-N...), but I am not clear on whether it's enabled by default. It seems like a good idea to have this, independent of current ARM vulnerability.

Out of curiosity, what are the benefits you're talking about?

Note that the demo sends your email address to the server if it's a hit.

$.ajax({ url: "/google_leak/save.php?info=manual_hit:" + email });

update: gone now. still pings that it ran. don't forget to hit ctrl-shift-r to bypass your cache.

However, there's might still be bias in who applies to speak at this kind of conference. They could have done more to reach out and encourage other groups to apply. I'd be interested to see the submission statistics.

They actually found the bug and released a fimware update fixing it before anyone's SSD was 5000 hours old, and SSD owners (at least back in those days) are generally expected to install firmware updates. It was quite well-publicised, too; they handled it really well, except for the actual release.

I wonder how well this would defeat microsoft asirra http://research.microsoft.com/asirra

The post is 3 years old. The issue was created today.

> I guess we've entered the age when browser choices have become moral decisions

For webdevelopers, they just as well may be. These older browsers can be a reason you can't use newer technologies like websockets or css animations, or have to use fallback libraries.

This seems to be a duplicate of http://news.ycombinator.com/item?id=4029676 , which is titled "German teen solves 300-year-old mathematical riddle posed by Sir Isaac Newton". It seems to be a German/Indian teen at a German university.

http://yow.eventer.com/events/1004/talks/1028 is another link to the talk, I think. I'm not sure though, the beginning is a bit different.