Deepseek recently published a paper in which they describe that vision tokens contain more information than text tokens and that this can be used to compress context.

We present DeepSeek-OCR as an initial investigation into the feasibility of compressing long contexts via optical 2D mapping.

Experiments show that when the number of text tokens is within 10 times that of vision tokens (i.e., a compression ratio < 10×), the model can achieve decoding (OCR) precision of 97%. Even at a compression ratio of 20×, the OCR accuracy still remains at about 60%. This shows considerable promise for research areas such as historical long-context compression and memory forgetting mechanisms in LLMs.

It reminds me of LLM caveman speak, it used to have another option to use Chinese instead of English. A language like Chinese is seemingly better at encoding information in fewer tokens and I think this is the same mechanism why OCR tokens work so well.

That said, I also doubt that voice messages are more efficient than text prompts, but it’s best not to waste too much time engaging with these sorts of LinkedIn posts (and LinkedIn in general).

The thing is that people don’t want to get that new game that that seems so fun to find out that it doesn’t actually work. Other games not working is seen as a sign of potential future trouble.

Great choice, using some of that VC money for things that actually benefit the world

I think the original post/image kinda was, but the response/retweet doesn’t seem like it

Fun fact the LessWrong guy Yudkowsky who was/is an influential person in silicon valley wrote one of the most well known Harry Potter fanfics. At some point he apparently tried to get JK Rowling’s permission to publish it.

Go seems like a good option to begin with; you can do a lot with it, and it’s not that complicated but does expose you to concepts like pointers.

There are plenty of very “serious” systems written in Go (e.g. Kubernetes), it’s not a toy language.

Some time ago I saw a video of a Pigeon trying to impress a Falcon, perhaps this could be the reason

Perhaps Kopia? It supports compression and deduplication.

It’s a field in the users database on your own device. You can change it yourself. If something stupid like that would happen you could bypass it entirely by just setting the field to something else.

Knowing about the world doping agency’s investigation makes it even better

‘Penisgate’ at the Olympics: why inject acid into your penis, and what are the health risks?

I’m not sure if they accept AI slop, but if you want to write a satirical paper you can send it with the information from their contact page

Well, then Google shouldn’t have just scraped the site then. It’s not JABDE’s responsibility to make their content suitable for LLM training

Indeed, that argument doesn’t really work. I suspect the argument will be that they’re untrustworthy and will give a distorted view of reality with subtle propaganda shown with a video of someone asking non open weights Chinese models about Tianamen Square or something.

Another approach is that they will form a cartel for running US inference focused datacenters and will pivot to selling services using it.

Maybe not as good as Claude, but they are good enough, and open-source, and free. The US market is going to learn the hard way why open-source curbstomps greedy bullshit.

Correct, the American frontier models Claude Opus, GPT 5.4, GPT 5.5, and Gemini 3 Pro still score better (while costing significantly more), but the runner ups are all Chinese models.

Then it’ll just come from Germany or France or elsewhere. It doesn’t take millions of dollars to train a good model, despite these US companies pretending that it does.

Well, it does. Deepseek-R1 cost $6 million and that was considered to be very cheap. Europe only really has Mistral’s models, Proton’s Lumo and several models that focus on transparency, ethically sourced training data, and supposedly better local language support (OpenEuroLLM, GPT-NL), but they’re by far not as good as other models and I don’t expect them to be for quite some time.

I’ve said this before. The Chinese models are significantly better and will outcompete the models from the US, it was just a matter of people realizing that.

My other prediction, being that they will lobby for tariffs or banning Chinese models outright also seems to be coming true.

A Dark-Money Campaign Is Paying Influencers to Frame Chinese AI as a Threat

Build American AI, a nonprofit linked to a super PAC bankrolled by executives at OpenAI and Andreessen Horowitz, is funding a campaign to spread pro-AI messaging and stoke fears about China.

I do wonder how Europe is going to react. Will they just focus on their home grown Mistral or will they consider Chinese open weight models? I feel like the EU is quite wary of anything Chinese and that many people won’t fully comprehend the actual security risk and that they will initially dismiss are avoid them, but they can’t ignore it forever. Qwen 3.6 35B which can be ran at home is already leaving Mistral’s latest models in the dust.

E.g. when you have a proprietary program that is only available on x86, but you want to run it on ARM.

It seems like you’re referring to dotfiles. You can manage these using a git bare repo.

We are obviously looking at things like Mythos, which is more sophisticated at finding vulnerabilities. In the next week or so, we will be changing our tack on coding the open and making our code public until we’re on top of that risk.

Most of our repos, unless they’re essential, will be removed for security reasons.

Security by obscurity because security vulnerabilities don’t exist if you can’t see them

Altman took the money and then OpenAI abandoned the non-profit structure to become a for-profit entity (2 years ago)

You could use something like Kopia and only include the files you want