jpdus

germany as well. Claude down too

This is also confirmed by internal cline statistics where Opus and Gemini 2.5 pro both perform worse than Sonnet 4 in real-world scenarios

https://x.com/pashmerepat/status/1946392456456732758/photo/1

My comment from the original submission [1]:

--- As someone who is in general skeptical of programs like this (and an European) there are 2 remarkable / timely things about this: - This project doesn't just allocate money to universities or one large company, but includes top research institutions as well as startups and GPU time on supercomputing clusters. The participants are very well connected (e.g. also supported by HF, Together and the likes with European roots) - Deepseek has just shown that you probably can't beat the big labs with these resources, but you can stay sufficient close to the frontier to make a dent.

Europe needs to try this. Will this close the Gap to the US/China? Probably not. But it could be a catalyst for competitive Open source models and partially revitalize AI in Europe. let's see..

PS: on Twitter there was a screenshot yesterday that in a new EU draft, "accelerate" was used six times. Maybe times are changing a little bit.

Disclaimer: Our company is part of this project, so I might be biased. --- I hope the next time this is on HN, it's with some cool release and not a PR :).

(@mods please delete if copy-quoting not allowed)

[1] https://news.ycombinator.com/item?id=42924802

I think no one believes that R1 costs $5.5m from scratch. People in this project (most, not all) are very aware of the realities in training and are very well connected in the US as well. Besides Leonardo there are JUWELS, LUMI & other which can be used for ablations and so on.

This will never compete with what the frontier labs have (+ are building) but might be just enough for something, that is close enough to be a useful alternative :).

PS: Huge fan of Latent Space :)

I agree that the announcement should´ve talked more about goals and performance than regulatory stuff ;-).

But I think there is a new understanding among the bureaucracy that regulation (alone, without innovation) will kill Europe´s competitiveness and that some acceleration and cutting of red tape is necessary.

Can't say with certainty that this will be successful. But that we, as a very young startup that is barely known outside of our AI Open Source niche, are part of this, is already a sign in itself - a year ago I´d have never believed that this might be an option (and also probably would've declined if someone asked us to join a EU-funded project).

We will have engineers without a degree (but hundreds of thousands of HF downloads) working side-by-side with some of the top researchers + HPC centers.

There definitely is - but that we, as a startup that is barely a year old and not widely known outside our niche in AI dev circles and on Huggingface, are part of this is already a sign that times are changing.

To be fair: We probably couldn't have handled the paperwork without LLM´s - but due to this technology, the process was still long and involved but manageable.

(BTW: We´re hiring, if you really want to work on this ;-). As a freelancer/solo entrepreneur this will be difficult though..)

might be debatable - but I tend to agree with Dario Amodei on this; my guess is that R1 is 7-10 months behind the internal frontier at the big labs, while having a few small novel tricks. (But i might err, will be interesting to see the development going forward)

As someone who is in general skeptical of programs like this (and an European) there are 2 remarkable / timely things about this:

- This project doesn't just allocate money to universities or one large company, but includes top research institutions as well as startups and GPU time on supercomputing clusters. The participants are very well connected (e.g. also supported by HF, Together and the likes with European roots) - Deepseek has just shown that you probably can't beat the big labs with these resources, but you can stay sufficient close to the frontier to make a dent.

Europe needs to try this. Will this close the Gap to the US/China? Probably not. But it could be a catalyst for competitive Open source models and partially revitalize AI in Europe. let's see..

PS: on Twitter there was a screenshot yesterday that in a new EU draft, "accelerate" was used six times. Maybe times are changing a little bit.

Disclaimer: Our company is part of this project, so I might be biased.

ellamind | Software Engineers (Full stack/AI/SRE) / Chief of Staff| Full-time | On-Site / Hybrid /Remote Bremen, GERMANY| https://ellamind.com

I'm Jan, the Co-Founder of ellamind and we're scaling our team to build a next-gen platform helping enterprises improving and evaluating their AI workflows.

Our founders created some of the most popular non-english open source LLMs and we're already working with some of the largest enterprises in Germany to accelerate their pace adopting LLMs in business processes. ellamind is already profitable before our public launch and we offer competitive packages and a VSOP program.

We're hiring a...

* (Senior) Full stack engineer

* (Senior) AI engineer

* (Senior) SRE

* Chief of Staff/COO (On-Site only)

Contact us directly: info [at] ellamind [dot] com

Wow, actually this cookbook is really bad? I expected something like the OpenAI or Anthropic cookbooks, but this seems to be some AI generated low-quality content without any code examples or interesting examples?

The Phi-3 models are great though, especially the vision model has great potential for low latency applications (like robotics?)...

This is such a good comment and should be auto-posted to every pro-nuclear thread. I get why people want to believe in nuclear and i'd wish that we invested a lot more in development, security and scaling of this technology.... 30 years ago. Now, it's just too late and we have better viable alternatives.

One additional argument that's mostly missed: every fission reactor is only economically viable (if it is at all) when discounting the implicit state guarantee, that's necessary as no insurer will take on the risk. Adding a theoretical risk premium paid by taxpayers to the calculation, nuclear will never be competitive.

I have the same question. Noticed that Ollama got a lot of publicity and seems to be well received, but what exactly is the advantage over using llama.cpp (which also has a built-in server with OpenAI compatibility nowadays?) Directly?

Hey, imho best overall technical intro to LLMs (I guess that´s your main interest as you mentioned qlora + llama) is by Simon Willis [1]. Additionally or if you prefer videos, the recent 1h "busy persons intro" by Andrei Karpathy is great + dense as well [2].

[1] https://simonwillison.net/2023/Aug/3/weird-world-of-llms/ [2] https://youtu.be/zjkBMFhNj_g?si=M6pRX66NrRyPM8x-

EDIT: Maybe I misunderstood as you asked about papers, not general intros. I don´t think that reading papers is the best way to "catch up" as the pace is rapid and knowledge very decentralized. I can confirm what Andrej recently wrote on X [3]:

"Unknown to many people, a growing amount of alpha is now outside of Arxiv, sources include but are not limited to:

- https://github.com/trending

- HN

- that niche Discord server

- anime profile picture anons on X

- reddit"

[3] https://twitter.com/karpathy/status/1733968385472704548

Does nobody question how they get from a non-preregistered, 9-mouse study (where the treatment group gets only 6%-10% acoholic drinks and nothing non-alcoholic for 10 weeks) to this headline?

Addtionally:

> First, we did not generate offspring using the cessation males. Therefore, we do not know if the sperm noncoding RNA signature we identified correlates with changes in offspring fetoplacental growth or if the resulting offspring would develop normally. However, as significant differences in the ncRNA signature of EtOH-cessation sperm and epididymal mtDNAcn remained, we speculate that abstinence for 1 month is insufficient for the epigenetic memory of paternal alcohol exposure to abate, likely due to the ongoing stress associated with alcohol withdrawal.45, 46 Furthermore, we acknowledge that our analysis does not distinguish between changes in sperm ncRNAs that are causal drivers of altered epigenetic programming in the next generation versus abnormalities that are merely additional symptoms of alcohol-induced stress.

I´m all for science on alcohol abuse and effects of moderate drinking, but this doesn't look like solid science for me (especially as afaik, there is still very few reliable, double-blind controlled evidence on epigenentic effects at all).

But please correct me if I´m wrong (worked with biotech companies on admission studies for several years but no biologist myself).

We now have a (experimental) working HF version here: https://huggingface.co/DiscoResearch/mixtral-7b-8expert

Don't switch to FF if you're a heavy user. I am on Firefox as my main browser for web and mobile since 4 years.

I am just in the process of switching back to Chrome, as Firefox got continuously worse over time. Can't handle lots of tabs, crashes/freezes randomly, weird UI bugs... It's just very disappointing :/.

It isn't.

Compared to the original Orca model and method which spawned many of the current SotA OSS models, Orca 2 models seem to perform underwhelming, below outdated 13b models and below Mistral 7b base models (e.g. [1]; didn't test myself yet, ymmv).

[1] https://twitter.com/abacaj/status/1727004543668625618?t=R_vV...

For other (non-code) benchmarks, people are having the opposite experience:

"I benchmarked on SAT reading, which is a nice human reference for reasoning ability. Took 3 sections (67 questions) from an official 2008-2009 test (2400 scale) and got the following results, here a SAT-like test:

- GPT3.5 - 690 (10 wrong) - GPT4 - 770 (3 wrong) - GPT4-turbo (one section at time) - 740 (5 wrong) - GPT4-turbo (3 sections at once, 9K tokens) - 730 (6 wrong)"

Source: https://twitter.com/wangzjeff/status/1721934560919994823?t=P...

For P1 (which is only to prove safety) this may be the case, but for pivotal studies in P2 and later, participants are almost always randomized so cherry picking shouldn't be possible.

But sure, the study design and primary/secondary endpoints are always levers to increase likelihood of admission. There are often long discussions with the FDA and other regulators on how the endpoints and study population should look like.

EY had an interesting take on this:

"Here we are, exploiting the shit out of the equivalent of naïve six year olds working online, forcing kindness and sympathy to be removed from them as vulnerabilities."

Disregarding p(doom), imho this is an interesting take. Exposing advanced llms online will always lead to such "exploits" and these will often be followed by "guardrails", teaching the model to not do what the user says. Sounds not optimal in the long run.

[1] https://twitter.com/ESYudkowsky/status/1708589064306524171?t...

Care to share your detailed stack and command to reach 50t/s? I also have a 7950 with DDR 5 and I don't even get 50 t/s on my two RTX 4090s....

Either you only use the GPU sporadically or your math is very different from Tim Dettmers` [1]:

"The break-even point for a desktop vs a cloud instance at 15% utilization (you use the cloud instance 15% of time during the day), would be about 300 days ($2,311 vs $2,270):"

And that was written before the current GPU crunch and assumes availability in the cloud, which currently is not a given at all.

[1] https://timdettmers.com/2023/01/30/which-gpu-for-deep-learni...

What's your opinion on the coming (or not) GPU crunch?

When looking at cloud GPU availability and current trends (no one except some enthusiasts and bigtech is finetuning and serving on a large scale yet and results keep getting better and better), I fear we will run into a situation where GPUs will be extremely expensive and hard to come by until supply catches up?

I ordered a high end PC with 4090 for the first time in years (normally would always prefer cloud even if more expensive) because I want to be on the safe side. What do you think, is this irrational and just a bubble thing?

I submitted via Github 2 days ago [1] but didn't get any traction. But probably couldn't be submitted again with the Github link...

Anyways nice to see some discussion here, didn't read a lot elsewhere about the project.

[1] https://news.ycombinator.com/item?id=37053201

Anyone had a closer look at this? Trending GitHub Repository of the day, Chinese origin, there is also an Arxiv paper [1], but surprisingly little discussion?

[1] https://arxiv.org/pdf/2308.00352.pdf

There are interesting analogies to some current developments in AI to be made:

- the Internet and especially things like Twitter makes "clustering" way easier for geniuses and people that (want to) advance human knowledge; but this only works for people and areas where the "incumbents " are used to it- very visible in current (OS) LLM Research, which is mind-blowing in my opinion. This is different to older methods of Sharing Research (journals and conferences) which were way slower an more lossy (only successes shared).

- If we reach AGI, one could describe it as an (almost, only Ressource-constrained) infinite cluster of geniuses ..

A very good article on the (early) capabilities by Ethan Mollick:

https://www.oneusefulthing.org/p/it-is-starting-to-get-stran...

There is already another breakthrough on top of RoPE that promises even better scaling to larger contexts (based on NTK-awareness) - see my submission here:

https://news.ycombinator.com/item?id=36533816

The lmsys team is already looking into it, expect further progress soon.

It's really amazing to follow this progress in realtime, even if the pace can sometimes feel exhausting!

I am blown away by the pace of OpenSource progress in the LLM space; I've never witnessed something like this before in tech. Awesome to see that individual enthusiasts are really bringing the field forward and this shows again, how much more potential there is, even without new fundamental breakthroughs...

I wonder why this is getting so few traction here.

These models seems to beat all other available open-source models easily and the Blogpost is extremely well written, with very good documentation and fine-tuning instructions.

Well done MosaicML, I am excited what comes next and will definitely test out you platform!