But what was the likelihood of this bug to be exploited by malicious actors?

I don't understand the question.

Somehow we encoded our human thinking or it learned it from all this training on user data.

Honestly right now it's mainly stagnation in frontiere model capabilities. Most of the recent afvancemdnts are towards generation speed, compression and tool usage. The quality of the models are not improving at the same rate as before. I doubt this big gap will continue, given that open source and especially chinese labs keep pushing well documented frontiere papers.

Iran was known to have such capabilities, it's baffling the US wasn't more prepared in its gulf bases.

> it's baffling the US wasn't more prepared in its gulf bases.

Probably want to drop the assumptions about it having anything much to do with US interests. Better to start looking at who has had the alliance that contained them damaged and their oil sanctions lifted.

If only there was a 4 year long war where thousands of drones are flown every day both on offense and defense that we could have learned from ..

Problem is that there was too much propaganda in that war, that parsing propaganda is too difficult even for military watchers, let alone general public. Only when american weapons are being destroyed that, US MIC is willing to acknowledge that may be million+ usd missiles are not solution to cheap drones.

Problem is also that your “Secretary of War” has fired two dozen of your most experienced military leaders since coming into office.

When the history of the American demise as a global superpower gets written, this war and the government behind it, will merit a beefy chapter.

https://www.bostonpoliticalreview.org/post/pete-hegseth-fire...

These traitors will eventually be all prosecuted. They are all traitors with putin connections, every one of them.

There will be no prosecutions. Even if there's a situation where Dems regain power, they don't have the political capital or efficacy to prosecute.

Like how assiduously Obama went after Bush Jr. administration.

...and how decisively Trump was prosecuted for the 6/1/21 attempted ~coup~ tourism, and for how thoroughly the Epstein child abuse ring was dismantled, and...

Yes, the only chance the US has going forward is to primary all current incumbents and hold both party leadership accountable for complicity in treason.

Even that won't matter. The problem isn't the elected officials, the problem is that most of the county doesn't care either way.

nobody will prosecute them, unless there is regime change in the USA

Haha, by whom? There are zero higher-ups who are actually getting institutional backing and are in favor of this.

Look at how Mamdani didn't even get any backing. Quite the opposite, he was obstructed. And he's 100x more palatable to them than the idea of prosecuting the traitors.

This is a completely unrelated problem, the US MIC is heavily incentivized to invent new problems.

> I got to say people also seem to be missing really simple tricks with RAG that help. Using longer chunks and appending the file path to the chunk makes a big difference. > > Having said that, generally agree that keyword searching via rg and using the folder structure is easier and better.

It depends on the task no? Codebase RAG for example has arguably a different setup than text search. I wonder how much the FS "native" embedding would help.

A stupid question, what's the risk?

The risk is minimal if you control or trust both networks. A network boundary is a natural choke point for access control, so that's where it's usually implemented. For an ipv4 boundary router (as is the topic of the post) you almost certainly need to configure Network Address Translation because your internal network addresses are non-routable on the Internet (at uni my dorm had public IP addresses for each student computer, fun times).

As for the GP's example, running VM's or containers* on your own machine? I'd say the default ACCEPT policy is fine. However, silently changing such a setting on software installation is a problem because if the machine is multi-homed (i.e. has more than one network interface), you've now created a network route outside of the network admin's control.

* The default for docker and podman is to use a private network, not a bridge anyway.

It's can also commonly be a problem if for example you are connected to multiple LANs via wireguard or similar.

Basically you're introducing a hole. For example, if you have some devices in your network (like a dodgy TV box) that are not supposed to reach the internet or other parts of the network, the computer with net.ipv4.ip_forward=1 could be used as a pivot. Depending on the routing tables you probably would also need to enable IP masquerading (NAT) to allow bidirectional communication.

that you'll get it wrong, I suppose.

after all, most routers/WAP/gateways that you buy today will have linux on the inside, configured similarly.

In almost all Linux based router setups: folks end up using 6to4 tunnels, packet marking, and interface routing priority.

Setting that up with safe/fair bandwidth-sharing requires intermediate IT skill level. Still a great hobby project =3

What did Apple do?

Grabbed up as much ram as they could, nearly no questions asked, at above market rates in some cases, ramping up the perceived demand and decreasing supply significantly.

There is a possibility the IRGC are trolling, given it's April's Fool today.

Kill they Supreme leader and 40 other leaders, destroy their Navy and Airforce and give them 30 days of B1 and B2 night and day bombings, and they decide it still worth it to joke on Aprils Fools ? :-) I have to give to them...

I guess this is supposed to be funny but I wouldn’t take that chance.

This is the bet of many of the big AI companies, and why they're subsidizing majorly the calls. With the latest cracks by the US gov, it seems Anthropic is starting to reduce those subsidies given their edge in the game. I am starting to consider local models more seriously beside just testing, but nowadays the ram/gpu market is bloated.

Local models just don't seem that useful for me for these particular tasks yet - the most recent versions of Codex and Claude Opus are the first time I've found them to be particularly useful in a "real engineering" context that isn't just vibe coding.

Google's TurboQuant might help address this, but it also might just widen the gap even further.

I am far on the skeptic edge when it comes to the generative AI side of ML tools though, so do take my opinion with that weight.

Turboquant is totally irrelevant compared to current quantization methods. It has been thoroughly test by people who build inferencing engines for local models. It's all talk no actual meat to it.

Do you have any reading on this? I find it hard to believe something announced a week ago has been “thoroughly tested”.

Their paper TurboQuant (TQ) is not new per say. It's released last year, and heavily rehash of old ideas that were released a year prior (RabitQ). There is also [a bit of drama](https://openreview.net/forum?id=tO3ASKZlok) there that boils down to what it seems a bit of malpractice for google's researchers. TQ does few things: it claims better compression quality and speed, and better KV cache handling. Currently KV cache takes a load of resources beside that of the model itself. Many people applied different quantization strategy for it, but the quality degradation is a too apparent. Enter Attention Rotation. This seems to have genuinely helped KV cache compression as per [llama.cpp latest tests](https://github.com/ggml-org/llama.cpp/pull/21038). On the other hand, [ik_llama.cpp](https://www.reddit.com/r/LocalLLaMA/comments/1s7nq6b/technic...) did tests on the quality of turboquant-3 compared to IQ4 quantized models, and yhe quality degradation is much worse. So it's 2 things: KV compression -> good. Turboquant quantazation -> not good.

This is pretty much my case right now. BM25 is so useful in many cases and having with with postgres is neat!