acje

I remember vaguely from a presentation by Yann LeCun "Intelligence is not what you know, but what you do when you don´t know". I find it helpful when trying to build an intuition for how to understand the LLM tool.

It is ok that you feel this way. As explicitly stated the report is _generated_ and id does require a different mindset to read AI generated reports. It contains quite a bit of information about attacker behavior and details of the timeline. It also linkes to a story that is now largely human narrated and I´ll keep improving on that as I find time. Having to juggle three separate issues yesterday did not leave much time for manual writeups.

I have generated a report based on my own exposure to this attack today. The report was updated as the attack unfolded; https://acje.github.io/systems/arrayref_incident_record/

I got to watch this attack unfold pretty much in real time as my agents worked the issue. Here is my writeup https://acje.github.io/systems/watching_a_supply_chain_attac...

Thanks for the good links. I think we generally have become so accustomed to the scaled up von Neumann strategy that we don´t see how much efficiency and performance we leave on the table by not building much smaller memory hierarchies.

Shameless plug here, where I explore possible gains in efficiency, performance and security by scaling out rather than up; (no subscription) https://anderscj.substack.com/p/liberal-democracies-needs-a-...

A resource like this is a good place to discuss where the two languages are near and far. Of course there are going to be styles within each language that differ as much as the languages themselves.

There is some elaboration in part four of the series. A fifth part on actor model, gaps and surfaces is in the works. Part four https://lnkd.in/dEVabpkN

I find it inspiring that we are getting to where we are dealing with models that classify vulnerabilities at a systems level. However I also think we are kind of barking up the wrong three. There is IMHO something wrong with the current strategy of scaling up the von Neumann architecture. It leads to fragile software partitioning, noisy neighbors and both slow and sometimes unintended communication through shared memory. I’ve tried to lay this out in detail here https://lnkd.in/dRNSYPWC

2028 human text is too ambiguous a data source to get to AGI. 2127 AGI figures out flying cars and fusion power.

There was this "end of history" idea where Europe, elites or not, believed that democracy would follow globalization and free trade. This turned out to be very naive but still common belief. Now Europe will have to fend for it self. In the long run I think this will benefit both US and Europe, but short term it will hurt both. Also I don´t think Europe will pivot away from US, we still understand that at least half of US shares our values, but US can´t be trusted to stand up for these values anymore. Same as Hungary, Turkey. I personally see it as a good opportunity to get some action going in Europe, as I discuss here in the context of building digital infrastructure: https://lnkd.in/dRNSYPWC

Shares aren’t the sole mechanism for influence though. In Russia there are open sixth floor windows one could fall out of. In China you could disappear to a camp for a few months. Shares are kind of soft in comparison.

My biggest fear is that Trumps deterrence of CRINK countries will cause XI to miscalculate. Other than that I think this is manageable. EU will get a boost as the internal awakening materializes. As an European I had difficulty understanding why Trump was even an alternative, but I have come to realize that the Plutocratic nature of the US was causing more suffering for the people than was easily observed from here.

Both are examples of communication by means of frequency modulated and amplitude modulated electromagnetic waves with distortion from a moving three. Also a good example that a large change in quantity is a change in kind. Probably a legit analogy imho.

So it is just buggy cache invalidation? I guess eBay will eventually fix it.

Also the seller is perhaps at fault for not naming it wisely?

Cache invalidation and naming stuff..

Funny. This is genuinely one of my deepest insights after 17 years in IT, some of it even as lead on infrastructure and platforms. Perhaps not well formulated, ok. And to date one of my most down voted comments on HN. What are the incentives supposed to achieve?

I generally view churn and nice primitives to be the essential balancing act for a platform. Perhaps any leading edge (software)system. We aren’t going to get anything complex perfect the first time, so it is change or mediocrity.

Has it been shown that the hype cycle is just Dunning Kruger effect for populations? I feel like every cycle starts with the experts standing on mount stupid confidently explaining how radical this new thing is going to be.

Pretty sure we can´t divide Mississippi by zero

Good question and I don’t have a good answer. My intuition is that democracies are more dependent on transparency and that lower complexity and higher security would enable more distribution in control.

I look forward to the decline in interest in generative ML. There is a screaming need for secure online services to enable democracies to face off the threat from authoritarian regimes. To do this we need hardware that enables actors as in actor model with truly private state, not sandboxed where there is an external entity that can observe its state. Today pretty much all designs has a perverse von Neumann architecture where state is shared across different compute devices like network controllers and management engines. And the software stack is more of the same sandboxing. Apple includes a Secure Enclave on its SoC where you may only communicate with it by sending messages like a proper actor, but why aren’t servers made entirely of secure enclaves? If the memory of each enclave was private by design in hardware it shouldn’t be necessary to encrypt it.

This has some analogies to how I interact with management as an architect… and with devs ducks

+1 for quitting. This can be life saving. Finding new purpose then becomes possible.

Yeah that sounds like looking for fairness and effectiveness in buffer management in a shared environment. I was thinking of measuring latency of x.

The M1 cards I mentioned was notorious with 1GB (?) buffer and a massively oversubscribed packet processor. Causing extreme jitter.

Having thought about this today and read some responses here, my best proposal would be to require endpoints available for the FCC to measure and then just create a service to show the results. That would give so many more vectors of improvement over time. How to measure, report such that users understand, capture change over time. No mater what number one puts on a label, that number will likely become misleading over time.

Nice videos Dave. I guess I have personally given up on effective buffer management. Perhaps if ipv6 and infiniband becomes the underlying infrastructure? There is just so many layers of abstraction hiding no longer useful decisions in the stack that I have just decided to leave infra and networking behind for a while to see if one can make a difference elsewhere.

Measuring latency under full load is like measuring how fast you can drive into a crossing and not making the turn. It is meaningless. See the fantastic explanation by Gil Tene in the link above.

With best case latency you can determine if the service can be suitable for real-time or not. There will always be buffer effects and they will vary by other user activity and the complexity or the packet processing. These effects will largely be unknowable in advance and the part that is knowable is extremely difficult to communicate to an average user. They don’t really get HDR histograms https://github.com/HdrHistogram/HdrHistogram.NET/blob/master...

Yes, that is a good one. I guess a TLDR for the FCC would be; a) No do not bother because most user will not be able to comprehend meaningful latency metrics. b) Report best case latency under idle load to nearest IXP because this would compose well with bandwidth properties when reasoning about the suitability of the service for different workloads.

Of course there be dragons. Oversubscription in various parts of the network can have interesting behaviour. Especially when coordinated user behaviour or complicated packet processing is at play. Think encapsulation or deep packet inspection. My worst case experience includes Cisco Nexus M1 32x10Gbps line cards maxing out at 1-2 Gbps throughput and 1+ sec worst case latency because of OTP. This is a datacenter core switch. And F5 WAF eating random packets because something looks like a VISA card number, causing a retransmit, that again shows up as high latency at higher levels.

“Gil Tenne - How not to measure latency” some of the best infotainment on YouTube. Sorry no time to find exact link. There are many.

I don’t think SMT makes much sense if you have enough cores to more than max out the power envelope of the socket when they all run at max frequency. You can then dynamically tune the system to get as close to the power envelope as possible with the available threads by adjusting the power usage of each core. Also the cores would be simpler. Meaning more power efficient and potentially eliminate some attack vectors as mentioned earlier in this thread.