himata4113

With normal training you're simply training for higher loss for things you don't want while keeping the loss low for things you do. Same would apply here except you would target low eval for behavior you do not want the model to have rather than trying to increase eval.

It doesn't really matter how accurate the post is, what it shows is what one could assume how you'd end up with a system such as RSI, how you could prune harmful capabilities from models and build a more efficient training system with biases including and not limited to emergent behavior in astra that might have come from something they've never even trained the model on regarless if ML engineers can already do similar things.

None of the previous studies applies because Astra simply did not exist and this is worthwhile NEW knowledge that has been publicized and it is very interesting to see how it can be compared with other models as both openai and anthropic are keeping any and all RSI development completely private. The work is groundbreaking in a way that there will be a lot more experiments of this nature which will contribute to increasing the public knowledge of such systems.

I was lucky enough to know that all I need is presence in my life. I don't really care for interaction or romance.

Remember that having children, getting married, living a fulfilling life is just the status quo and not the path your life has take. Be or do whatever makes you happy as long as you're not harming others.

Might be a bit late, but it's all about the ride and leaving something behind for others to enjoy is a nice extra.

End result is the same though? Cause vs effect.

vendor oem's shipping broken firmware, enterprise customers complain to ubuntu, ubuntu fixes it in the kernel driver

any broadcom networking hardware, on debian it will constantly drop out @10Gbps link speed on copper is one such case that I can describe. But over the years there's been plenty of related issues with e810's and whatnot. Disks, disks will also stall or drop out which doesn't happen on ubuntu kernels.

I personally only rely on 3rd party libraries that are complete. I will never use a library that gets consistant feature updates. I do make an exception for openssl rusttls and whatnot since they have to update by design.

Their kernel is vastly superior when dealing with buggy enterprise hardware.

You don't have to keep paying AI subscriptions one the product is finished.

There's a lot of very weird websites. You still benefit from checksumming your own assets to prevent cache poisoning and so on.

I'm fairly sure the ZDR on their site impled that neither the prompts or the model output was collected before they've changed their site few dozen times.

They still mention: "meaning they do not store or train on your data", but only on the providers side which makes quite a lot of sense, there's very little reason for them to store anything except some analytics data in cases like these.

usually scripts have versions, this protects you from compromised cdns as well where you publish assets in a different location compared to where your frontend is served.

If a library serves a script and your site manages user funds you should probably avoid using it or use a cdn that does.

I am not quite sure, but there's also other integrity mechanisms that gurantee that a file was signed by the library developer via a served header with the signature of the file.

I like this analogy that I've come up with: Non-writers can produce impressive articles that are unbearable to read the same way non-programmers can produce impressive software that is unbearable to use.

This is just the beginning, once kimi k3.1 is out with cyber curated dataset post training we will likely see far more advanced attacks.

For now most of these appear to be related to supply chain attacks which can be avoided with basic discipline and using features that browsers had for years: https://developer.mozilla.org/en-US/docs/Web/Security/Defens...

low attack surface applications is the future.

No? Cryptocurrency has a ton of utility: it simplified ransomware attacks, enabled the deployment of supplychain attacks to steal funds from unsuspecting users, made illegal gambling accessible to everyone, memecoins, efts and safe storage for any stolen or illicidly gained funds until they could be laundered and the list goes on likely exceeding the character limit.

The biggest benefit is that you can only withdraw what you need into the real world, if you gain millions of dollars you can life off that money for the rest of your life with tiny withdrawals of $2000~$5000 or so serving as a stable stream of income which allows you to have risk free investment into real-world activites that would normally strain your finances and induce sunk cost fallacy.

Losing $5000 doesn't mean anything when you know that you will have $5000 tomorrow.

North korea has enough cars for a traffic jam so that's definitely a win for the people's democratic republic of Korea, the only korea to exist.

depends on how they're spliced, they have to be tuned for each configuration as there is no silver bullet. It also requires customizations to sglang/vllm to get it working in the first place so it's definitely pretty involved. However, it's still cheaper than obtaining more gpus.

It's lucky for everyone that this happened to binance, other exchanges would have likely caused a lot of issues for people holding money in said exchanges.

But it's also surprising that it happened to binance at all. This checks a lot of the boxes for being, at least indirectly, an inside job.

use MoE offloading

After posting a few questions to SO I often ended up not getting a response or solving the problem myself after further research which in reality sums up the site is a whole.

1) If you already have enough programming experience to understand algorithms and abstract concepts in programming you can find the answer by doing research.

2) More often than not the answers were citing answers that already exist elsewhere.

3) A lot of the trivial answers were marked as low quality, dupes and so on when these were the people who could benefit from SO the most.

Eventually I stopped visiting the site around 2 years after getting into programming since it rarely provided anything of value.

--

Serverfault was great and I will miss it.

no, but as with most statistics it is a bell curve which means that this generalization is generally okay.

being smart is also being lucky, half the population is below average.

Indeed, even though I can afford quite a lot I keep my monthly expenses below a nice 1000. I have to expend mental effort to justify anything above that and it has helped me a lot when dealing with a sense of meaninglessness as lifestyle creep will simply lead you to chasing for more.

Best things in life are free, or at minimum, very cheap.

It's ultimately a choice, work hard and earn a lot so you can enjoy early retirement or grind out a balanced life until the day you die.

It's a balance of risk and reward, you might work hard and never get lucky enough to retire early, but more recently we've seen how a balanced life can backfire @ FAANG layoffs.

You are indeed correct, but also you can use a slightly less jailbroken model to build a stronger model, this is very effective on models that are of the same checkpoint as they're more likely to "feel" what makes them refuse.

This seems to be a hit and a miss, there's no visual indication that opus 5.5 is better even though in 9 tests out of 10 it will outperform astra especially when it comes to niche domains and will absolutely destroy them in anything requiring raw intelligence rather than a naive search, naive search is what makes openai models extremely capable in scientific fields though.

The score hit on smaller models is a welcome change especially when compared to models that nearly 10 times larger.

Grok being punished quite a lot is also a very welcome change as the model is frankly not that good.

I would love to see few more points knocked off haiku by introducing steerability/creativity benchmarks.

vouch this comment since this is a pretty common question people have, I don't believe it should be dead.

This is entirely self inflicted as SGLang is optimized for (near) maximum performance. Another issue is that this is dealing with a device driver rather than system memory which rust has very little control over, as in there is no native implementation in rust that allows you to address multiple-memory buffers. Think how old games had 16 different memory blocks all starting at 0 since the cpu could only address 65 kilobytes.