> Wonder what it means for safety and interpretability at this level of availability.
We already know from official model cards that the GPT-6 models are substantially less monitorable than other (non-looped) models. Which is not surprising, since they are doing some amount of Neuralese thinking between each token.
LLMs don't produce Lean proofs directly, they produce the proofs in natural language, and only translate them afterward. The fact that we have Lean has nothing to do with LLMs being good at theorem proving.
(I think we mostly agree.) From a previos post about Navier-Stokes, the English version may have small holes and the AI make a few changes and fixups when translating to Lean.
This is simply not true that they only invoke lean at the end. They invoke it multiple times in between to reduce the search space like GP said. See how e.g the prove2me harness (which anthropic used for FLT) works for more details.
Small errors in proofs can often be fixed easily. So it's not that the whole thing would be useless. Also, he has a Lean proof, which rules out these small errors.
And it will not be limited to LLMs. In just a few years, they will likely figure out robotics and continual learning. Then all jobs are affected. I'm very worried about what's coming for all people who rely on wages to survive.
All jobs will be affected before robotics. When there are less white collar jobs there will be more people competing for blue collar jobs.
More workers competing means lower pay. Less good paying jobs means less customers. Less customers means less jobs. This is an over simplification, but is one possible future.
On a message board institutionally subservient to Capital there is all the talk about how LLMs will take wage jobs away, but little analysis of why people rely on wage jobs to begin with.
But never mind that. Get money or die. Every rat for herself.
It's unclear what "PE" means and whether it really indicates that investors expect a collapse in profits. Even if they do, you are making predictions based on investor predictions, which is a two-step prediction. It would be better to first find out why the investors are predicting a decline in profits (if indeed they do) and only then use that as a basis for further predictions.
SSD and HDD prices going up always seemed strange to me, because LLMs don't need huge amounts of either. So the AI induced increase in demand can't be that large. What LLMs do need is large amounts of DRAM.
So it would be more interesting to predict when DRAM price could go down again. Several new fabs are planned to come online in 2027/2028/2029, but it seems unclear whether or not the substantially increased supply will be absorbed by further growth in DRAM demand from AI companies.
IMO mi think its probably for surveillance capitalism purposes. the ai companies wanna keep data on how u use ai, and other companies now wanna run all ur data thru ai, so its probably an exponential growth of data surrounding the ai boom
Because it is a user choice, and there are applications where it is inappropriate. For desktop Linux it should probably be on, as long as its clear to the user - but yeah, for many other Linux use cases its inappropriate.
Maybe it should be generally on for desktop-focused distros and generally off for server-focused ones. The distros that focus on doing both well might want to recommend it in the installer when told it's primarily a desktop installation and explain the tradeoffs.
For a server, it really depends on what kind of service you're providing whether it would be helpful. We might want to let server admins, SREs, production engineers, or platform engineers to make an active decision about it.
In most cases you'll really want zswap rather than zram anyway.
There are plenty of scenarios in which servers are run with only nominal swap sizes or none at all and systems don't run out of RAM.
Let me size my hypervisor cluster to demand and let the system decide where to place the guests. If the hypervisors are crashing due to full RAM, that's down to planning or budget issues and not a reason to put additional unpredictable latency underneath the guests.
Let me size my SDI to UDP TS video streaming servers to the number and size of streams they'll service. There's not much in their memory pages that's going to ever be idle enough to compress, so why boot the kernel configured to do that? It's just risking extra packet IAT for buying not much benefit.
But in general, yes, zswap would be a good thing compared to swapping to a drive, unless you are mostly swapping things that are really truly idle a long time. Default swappiness on a lot of distros is around 60 by the way, and on servers it's often set closer to 10 or 15.
We already know from official model cards that the GPT-6 models are substantially less monitorable than other (non-looped) models. Which is not surprising, since they are doing some amount of Neuralese thinking between each token.
reply