xiphias2

Mojo may be interesting to me, but they should really split it from Max, which is some closed stuff I'm not interested in.

They should show some performance comparisions between PyTorch and Mojo, PyTorch+kernel compilation + Triton vs Mojo, ThunderKittens vs Mojo.

Exploitbench in itself isn’t that different from gain of function research in computer code, I can imagine the next being ExploitAndCopyYourselfBench.

Sure, and spent most of it experimenting with something nobody knew would work in production or not, and needed a lots of breakthrough ideas that are now known as the machines exist.

Not really, the only reason I (or any other programmer) haven't ever hacked into a system to make my life easier (not to do bad things) is because it's illegal.

It was always easier than making a system secure.

Everybody who has rewritten software understands the current phase: it mostly works, but there are always things to fix to make sure that there are no regressions, and the pressure is huge for any release.

I still believe it was the good decision, but I also know that I wouldn't be the first person to run the release in prod.

I think Jarred should start making release candidates instead of releases to take some of the pressure off.

,,You can play this game of whack-a-mole indefinitely if the state of the system is concealed''

Not really as one of the main goals ofr ARC-AGI 3 was measuring task efficiency on unseen games.

I'm sure there are cheats everywhere but the most sensible thing is to just accept that the LLMs of today are much more intelligent in solving reasoning tasks than the ones from half year ago.

My own private benchmark shows the same thing.

So far it seems that he's winning as OpenAI is being sued, not Liu

You need open hardware and open software at that point and you won't be able to use government identification as they depend on closed source parts of the Android ecosystem. Also you need identification for side loading apps at some point.

Non of these laws stop you from opting out of surveillance, but altogether it gets so hard that at some point you get more suspicious and tracked if you do all this than if you don't do any of these.

It's because there's no real solution to the alignment problem.

Humanity is on aggregate optimizing for increasing collective intelligence and I don't see that stopping as the main goal when AI gets smarter than humanity, just hope that we all will have a nice human life.

If you want something real, look at Clean Code Horrible Performance youtube video if you haven't.

I also found Bob Martin entertaining, but after trying to follow his suggestions I just got into the trap that Casey Muratori shows: even if I don't care about the performance of the code at the start, the amount of abstractions Uncle Bob advises makes it just too easy to make the code extremely hard to optimize later.

https://youtu.be/tD5NrevFtbU?is=vuVfjbsINQrtfvbC

EU needs to decide if it wants to do data processing or not.

If it’s a yes, it needs datacenters and get a lot more energy.

If no, it needs to transfer data to US for training/inferencing on it.

But Micron didn't restrict the output: it stopped producing DDR2-3-4 completely, so it's not profiting from DDR customers at all.

,,The plaintiffs claimed the three companies reduced D-RAM supply under the pretext of transitioning to high-bandwidth memory (HBM). "The D-RAM oligopoly companies systematically coordinated the shift to HBM and the discontinuation of DDR3 and DDR4," they said. They added that Apple's recent sweeping product price increases were the trigger for the lawsuit.''

How can they do price fixing and discontinuing a product at the same time? It just looks like some companies are angry that AI / VC industry is outpricing them.

I think the sim cards are more important: he wrote that Nest switched to local recording mode and the police took the evidence.

ASLR is important, but you should be able to provide the random seed (I don't know if it's possible or not)

It was always the same, Google even lost a big lawsuit because it went too far in doing what the Biden administration was asking.

Twitter and Facebook also did what they ,,had to''.

The thing that's new here is that Antropic's growth rate was so enormous that Dario didn't have time to learn to lobby.

What's interesting is that all big LLM providers figured out that the revenue comes from coding as it can be trained with RL rewards.

It means that most people will never understand how fast the landscape is moving: non-coding and non-homework use cases didn't change that much since last year.

They needed to raise money for the purchase, they just did it (raising from public market)

I guess they will upload it later, it seems like an honest mistake to me.

Anyways SwiTransformer paper looks interesting and doing a post training to optimize for it looks interesting as well.

From what I understand Amazon's lawyer was telling the CEO that it's better to report a vulnerability of a potential cyber weapon (jailbreak) to protect Amazon from liability.

And then US chamber of commerce protected itself from liability as well to take export control law seriously after Antropic's CEO categorized Mythos without the cyber query filter as potential cyber weapon.

The problem here is that there's no real technical way to protect against the queries, just making them more expensive to create, as the models are getting smarter.

Also as the main jailbreak ,,technique'' is splitting the task into subtasks, and the main moat of Mythos is that it can solve more complex tasks, I also wouldn't categorize this jailbreak as serious.

This should be at the top: they uploaded the wrong model, they fixed it

Unlike CZ who allowed money laundering to happen, SBF directly stole from many millions of voters.

Trump looks smart enough to understand the difference in how much political power he would lose if he would pardon SBF.

It makes sense, he stopped contributing to Tesla significantly long time ago (he redirected the NVIDIA GPUs that were supposed to go for Tesla self driving to X.ai).

He wants full control back.

It's very strange web site, something is off with the math, $9B cost of chemotherapy went away too fast.

It's not sabotaging it by using a worse model but by changing your prompt in your background, which means it silently destroys your code.

Also I asked questions about whether it's safe for me for example to work on just compilers or just inference kernel optimizations and it refused to answer me.

If I can't even ask what I can do safely without my code being destroyed, I just can't trust it not to sabotage my work ever.

Of course they are afraid of it, haven't you seen Dario being angry of Chinese companies paying for Claude access (tokens = test cases) and training their own model from those?

Not for machine learning, just for security bug finding and biology

Cloudflare as well

I don’t understand one part of the licensing here: if it was just a license, can’t they relicense the software and hardware of LPU3 to AMD? Or hire new software and hardware people?

The new designs were their main asset besides the amazing talent that went to NVIDIA, not the remaining DCs.

The problem is the we couldn’t really tell part. Changes made to mature finished projects should be minimal and readable and understandable by humans.

Also rsync is handling copying binary data, it’s a project that’s super sensitive to hardware faults for example, which means it’s not just enough for the tests to pass.