Pinned Tweet
"I have no experience writing Windows drivers. Now I wrote one and without my driver everything works fine but with my driver Windows crashes all the time and my driver is on the callstack. Can you debug Windows for me? It crashes and I don't think my driver causes that"
#KillMe
Us then bombs Malaysia and this is how the war between men and machines will begin.
China puts data centers in Malaysia to get access to compute they can't get in China due to sanctions. The US puts data centers in Malaysia because every state bans data centers. AGI happens. Malaysia nationalizes all the data centers and becomes the sole hegemon of the world.
5
543
MMX was great thing at the time, and SSE and AVX later on was even better.
Microsoft should have spend a lot more effort making its software faster, smaller and more reliable.
This mixes Knights Corner and Knights Landing which had different architecture, different cores too and so a lot of things here are not correct.
Commercial success was also affected by different factors.
I worked on KNL bring up.
In the early 2010s, Intel's Xeon Phi (a descendant of their Larabee x86 GPU project) was a 57-72 core CPU coprocessor with an early form of AVX-512.
This is often branded today as a failure and a proof that massively parallel CPUs are destined to fail. Except AMD is now selling their 256 core Epyc 9996 and doesn't seem to be having such trouble.
What's the difference? Part of it is software and timing. We've just gotten much better at building software that can make use of so many cores. Telling people to write code that uses dozens of cores was a lot scarier to people 15 years ago than it is today. It wasn't fully understood how to do meaningful parallelism safely, and there was a lot more skepticism that it could be done at all. They also couldn't just ask Claude to do it.
Another part of it though was just bad architecture design; Intel at the time was very fond of connecting cores via a "ring bus". All cores are connected in a ring, and if one core wants to talk to another for any reason, it must send a packet that walks all the way around the ring until it lands at the desired core. The memory controller may have its own stop on the ring.
For 4-8 cores this isn't so bad, but it scales very poorly with dozens of cores. If your ring bus has 3 lanes and the average packet takes 20 hops to reach its destination, you hit 100% congestion the moment your code averages even 0.15 packets per clock cycle. In practice, you can still get bad congestion problems even well below this rate. Not to mention, dozens of hops for common memory operations is slow, and Xeon Phi was an in-order core with a very minimal ability to absorb cache miss latency. Time spent waiting on very slow and congested memory was time a core often could not spend computing anything else.
This isn't a proof that high CPU core counts are a bad idea. CPUs have since switched to various forms of mesh and tree topologies, solving the problem. These other topologies are trickier to build correctly, ring buses were legitimately simpler to get right, but they don't scale.
If you running jailbroken Chinese model locally on your own computer system, what is the risk to the defender?
We’re going to be in an insane world soon where attackers have access to capable models, while defenders (even ones *verified* through Daybreak / CVP) get refusals on patch analysis, bug finding, triage, tool building, verification, PoC development etc and cannot do their jobs (unless of course they also start using abliterated Chinese models, which is obviously a whole other set of risks!)
2
7
789
Is this the reason 四 is unlucky in Chinese? Do they know something we don't?
Has anyone considered using 4️⃣ as their private key? In unicode it's 7 bytes (56 bits).
This feels like a fair compromise between the two main stakeholder positions, "Private keys should be at least a little bit secure" vs "I want my private key to be 4"
2
9
668
PSA: Always use both versions and Windows memory test (on Windows of course). I have seen broken memory when one passes but not the other.
When you work for a chipmaker, let's just say you have to run memory checks every now and then... ;)
Finally the opensource version of memtest86 supports Arm.
Passmark's proprietary memtest86 is still slightly superior in it's post test evaluations, especially for DDR5 but for a decent and reliable test the opensource memtest86 is still really good.
phoronix.com/news/memtest86-…
Oh my God! Science has to figure out something unknown?! Nooooooooo!!!!!
Just for reference: no, that isn't my issue.
My issue is the many idiots framing the situation with this dumb "trampling the norms" narrative.
While, in fact, we're facing major methodological and economic unknowns on how to proceed forward with science.
9
676
Performance is determined by the compiler, not the language. Crappy C compiler will loose to great Rust compiler and vice versa. Compile, measure, inspect assembly, understand and measure. Experiment.
Maybe you can beat this compiler, or maybe it is not worth it.
It is entirely possible to have great compiler and not so great programmer who chose bad was to locate the data and cannot take advantage of space locality. Rewrite might choose a container that C just doesn't have and take full advantage of cache. Result will be improved perf.
7
328
That's like one chip these days, right?
US arrests tech CEO accused of smuggling $300M in Nvidia chips into China arstechnica.com/tech-policy/…
3
366
It tells us that a lot of people have "negative knowledge" - they need to learn something first to know nothing at all.
Or, in a simpler terms, whatever they know is just not true.
;)
I’m hearing this a lot lately. “AI alone is already better than AI + doctor.” (Last heard here @x @FutureJurvetson). If true, what do you think this tells us?
7
572
And by "check the logs" we mean we ask AI agent to check the logs and produce a report which we don't check but we have another AI agent which reads the reports and sends us a summarize email which we occasionally ask email AI to summarize if we don't forget.
3
16
781



















