fancyfredbot

What happens if the memory is freed in a different order to allocation? Does the language disallow that? Or does it get allowed but mean the stack can't shrink?

How are you deciding how many stacks to use?

Sad that that a "buy everything" strategy performs best. It's not a strategy which varies much based on other players behaviour which makes it hard to counter and therefore boring as everyone has to play that way. I guess where there's still room for adaptation is the auctions if the player landing on the space can't afford to buy. The bidding strategy here forces some player interaction and room for different strategies...

Mimo is very good.

I thought it was 6 Sol which was a rebadged Terra? You think 6.1 Sol is too? It does seem to benchmark better than 5.6 Terra? Regardless agree both models have been very bad for openAI's Sol brand. To be fair it's at least cheap compared to other OpenAI models.

Cool! XLA is pretty good + potentially a better compiler infrastructure than inductor which will lead to improved performance on all hardware.

Also I guess this is the final nail in the coffin of Tensorflow.

There's something really cool about this but also something sad.

There was a time when CERN would have been establishing partnerships with US labs. The US is not a CERN member but has provided a lot of funding in the past despite it's observer status. These days funding is getting cut, the US is not as trusted as a partner anymore and a European partner is seen as important. This is sad.

That said I'm happy for Mistral and I'm sure they'll be a great partner for CERN.

The UK spent 100m GBP on a 900m tunnel. Rather than protecting cars from rocks, this one protects trains from bats (or vice versa really) so structural requirements are presumably less stringent. It's weirdly pretty though.

I think you know how to do this. You are head of engineering at an AI startup. You know how to create and revoke API keys and how to share them with your interview candidates.

At that price I think you'd pick Opus 5.5.

It's perhaps a face saving thing to pretend you'll charge the same. Let's see if it actually happens.

The thing about Zuckerberg is that although he's deeply odd he has no ability to conceal this.

Sam Altman is much more dangerous because he is occasionally able to come across as human.

Of course you need talent but supply seems reasonably high. We have openAI, anthropic, xAI, meta, Xiaomi, deepseek, Alibaba, z.ai, Mistral, reflection, and Google all producing exceptional models.

The biggest differentiating factor seems to be compute. I know most Chinese labs would really like more compute. Google and xAI too, but largely because they sold compute to their competitors.

The point of the knowledge is that it helps you ask the right questions, not so much the ability to answer them.

Knowledge has been available just by asking Google for decades now. The LLM makes it easier but it's a difference of degree not of kind

Until the models are 100% reliable knowledge will be required in order to quickly spot issues and work efficiently with the model to address them.

Tenured mathematicians care about publication. My impression is that this is because they believe in the process of peer review. However they also have other incentives as their publication record affects their ability to attract research grants. They also typically share authorship with untenured students and postdocs who need publications for their own Tenure tracks.

You are right that most work gets published in mid tier journals. These are still peer reviewed. The difference is that they have a lower bar for impact. The bar for correctness is still very very high however.

I picked Nature because it's the best known academic Journal. I'm aware it's not a mathematical journal. I don't think that undermines my point since mathematical journals obviously exist although they lack name recognition with a general audience.

I am pretty sure that journal publication is the norm. I do know mathematics although I am not one myself. The mathematicans I know want/need journal publications as unpublished research isn't counted in their research output. I know Arxiv is typically used for preprints while awaiting review but publication is always the aim.

Don't take it from me though. zbMath did a study finding 60-70 percent of arxiv papers on math end up in peer reviewed journals. https://ems.press/content/serial-article-files/47210.

I think you are implying that my suggestion is unreasonable and exceeds the standards applied to science produced by "normal" human processes.

In fact we already follow exactly this process for human papers and have done for a very long time. Publications without peer review are treated with great suspicion. This is how we end up with journals of varying levels of prestige and rigour.

The process isn't flawless and there are huge problems with retractions, as well as weird financial incentives and rent extraction but there is definitely an increased level of trust in a paper published in Nature.

When You See One Cockroach, There's Probably More.

This can and should erode our trust in every single proof OpenAI published. The model is clearly faliable despite the lean proof, and clearly the output was't actually checked properly before release. Once these proofs are peer reviewed and published in a journal we might be able to trust them again but until then they are just slop, sadly.

I think OpenAI actually did the right thing by sharing everything with the whole community right now but I also hope that some significant credit will now go to the reviewers who confirm these 'proofs" actually work.

I'm usually very pro AI and the idea that people could vibe code their own games which combine exactly the things they enjoy appeals to me so was quite excited for this.

I played a few editors picks. "Westbound Frontier" is a kind of RTS. On mobile all the windows open on top of eachother and the text is very hard to read. The amount of resource required to build the first building after town hall was way too high by and you had to wait ages to do anything. There was an infuriating AI sound track singing "ooo-wee-ooo". 0/10 from me, broken and not fun.

Next I tried volcano run, a 3D infinite runner. Not my kind of game but a nice aesthetic and it actually worked. After about 30 seconds though the graphics changed to a really low contrast mode (night?) and I couldn't see the rocks I was supposed to dodge. 2/10.

Honestly genuinely a bit shocked Google launched this. It's obviously not ready, the "Editor" who picked these games should have known that. I am not sure if this is some kind of set of skills for Gemini or a fine tune or just a prompt but it certainly feels very low effort and embarrassing.

Pithy, but only really applicable to Amazon. I'll give you some tips for future attacks.

I'm not actually sure Tesla is a great company from any perspective so should be easy to find another sick burn for them.

Google is going to be a little more difficult. Maybe say something nasty about advertising? Or go for the monopoly angle.

The US has made sure that Mistral has a large market in the EU by temporarily preventing non citizens from accessing Fable.

A lot of European companies now want a model the US can't cut off, but also lack trust in Chinese models.

Some of these will self host Mistral but most will pay them by the token. It's not going to be a huge market or a huge margin within that market but probably it'll be enough.

Well, with this and Beam people are going to have to stop saying that western open models are dead. This is great news. Anthropic and OpenAI may have a bit more knowledge, talent and compute but they don't have a monopoly.

Investors looking for them to make monopoly profits are going to be disappointed The premium they can extract from consumers for their models will be capped. Tokens are likely to remain close to the cost of compute, a cost which is high but falling fast.

Also, yay Europe! Although the comparison between this and mimo 2.6 is not flattering...

You are Google, training a new model. You regularly benchmark the checkpoints. It's gradually improving as you train. When do you stop training and release the model?

You certainly don't do it when you are just behind the frontier. Instead you wait until you are ahead on a lot of benchmarks.

So, if Google release a new frontier model, the chances it's ahead on most benchmarks are statistically about 100% because if it isn't yet then they won't release it.

It looks gorgeous, and you can zoom in to an individual gate level and see lines switch from high to low

The problem is the rendering is wrong. Gates which are connected to two distinct inputs according to the text description which pops up when you click on it are rendered as having both inputs connected to the same wire. It's a bit of a shame.

The article is AI slop but impressively the author has managed to convince the AI to make a ridiculous double counting error which functions as excellent rage bait. I expect a dozen similar comments to have appeared while I type this.

I've got bad news for you on the whole humans dying thing. Seems like actually literally everyone is going to die. Turns out that this has actually been happening for a really long time - possibly since day one - and we've got no idea how to fix it.

I'm surprised you feel like you know muse spark 1.3 performance well enough to question the validity of the index based on this benchmark result.

Muse spark 1.3 was only released yesterday.

Usable certainly. But my impression is that useful models still need a bit more than 16GB. Something like Qwen 3.8 27B is useful but squeezing it into 16GB requires fairly aggressive quantisation which will make it unreliable (e.g it'll get stuck in loops) and won't leave enough space for a long context (which qwen 3.8 really likes)

What Jalapeño has done is pair huge bandwidth (almost as much as a full nvidia rubin) with a less powerful processing core.

This means that even during the inference stage, when most architectures are bandwidth bound, they are compute bound instead.

This is why they aren't using multi token prediction, as MTP only helps you when you are bandwidth bound.

The power saving they get is likely from a much lower clock.

Yes

I kind of want to blanket recommend but that's not very helpful

I would suggest starting with Andrej Karpathy's YouTube video: https://youtu.be/kCc8FmEb1nY?is=oiDsrBYJg_MUUmoD

This video is excellent. I'm a huge fan. Also the video is zero commitment and instantly available which makes it a good way to check you are interested.

The book by Sebastian Raschka is slightly less accessible but very reasonably priced and the experience of working through a book is a lot nicer than skipping back and forth in a video (for me). Sebastian's blog posts on recent architectures are absolutely great too.

I'm (more than) twice that age, but I've spent time learning this exactly this from videos by Andrej Karparthy and from books by Sebastian Raschka.

I didn't do it because it was useful to me in a practical sense. It's because LLMs are fascinating and I want to know how they work. From that perspective it's been a great experience. I have afirm grasp of the basics. This makes it much easier to understand frontier concepts like compressed latent attention. I can follow the field and understand it.

Not sure I would have got as much out of it at seventeen. I have a lot of background and experience which made it much easier to learn. I wasn't struggling with the linear algebra or with python. I already knew pytorch and neural networks. That helped a lot and I covered these tutorials fast and could skip over large sections. A few evenings and the odd weekend day over a couple of months was enough for me.

For seventeen year olds the tutorials are good enough to make it possible to learn this but it would have taken a lot longer to understand. On the other hand I would have learned a lot more. I think I would have learned a lot of valuable stuff.

However I also think 17 year old me was studying for his A levels and probably this was right choice in terms of maximising future opportunities. I'm not sure I think learning about LLMs instead is sensible. Indeed it might be bad advice. But I can absolutely agree with the sentiment.I think 17 year old me would have wanted to do this too.