smaddox

though we did not delve into the observation

Oh, the irony.

Gold/M2SL(Billion USD) is currently around 0.12. In 1980, it peaked around 0.45. Monthly average since 1960 is 0.11. In late 2011 it peaked around 0.18.

Gold / Global M2 would be a better metric, but I haven't analyzed that yet.

I switched to Windsurf.ai when cursor broke for me. Seems about the same but less buggy. Haven't used it in the last couple weeks, though, so YMMV.

For those who are actually interested in this field, the proper way to measure this would be with a four point probe. You do need a constant current source and a high-impedence voltage meter, though.

Also, you don't need to solder wires to the sample. But if you want to measure the hall resistance of a thin film of a semiconductor, you can solder a glob of indium on to four corners of a 1 cm x 1 cm wafer, put it in a magnetic field, and then do basically the same measurement as four point probe, except not inline.

You can train that size of a model on ~1 billion tokens in ~3 minutes on a rented 8xH100 80GB node (~$9/hr on Lambda Labs, RunPod io, etc.) using the NanoGPT speed run repo: https://github.com/KellerJordan/modded-nanogpt

For that short of a run, you'll spend more time waiting for the node to come up, downloading the dataset, and compiling the model, though.

That application was submitted in March 2024 and is on track for approval in December 2026

Huh? Is this something where there's multiple incremental steps in the process, and that date is just the final approval stamp, or does it actually just take more than 1.5 years?

Agree more levels seems like it could be beneficial. And another Meta paper published a day later shows how that might work: https://ai.meta.com/research/publications/large-concept-mode...

Oh, interesting. 120db is much better than what I thought possible with CMOS sensors. Thats competetive with the human eye.

Well, that definitely changes my opinion on how feasible/competitive camera-only autonomous driving can be.

Apparently they're using this CMOS sensor: https://www.onsemi.com/products/sensors/image-sensors/ar0136...

It's not an event camera, so it's very much taking images, which are then being processed by computer vision algorithms.

Event cameras seem more viable than CMOS sensors for autonomous vehicle applications in the absence of LIDAR. CMOS dynamic range and response isn't as good as the human eye. LIDAR+CMOS is considerably better in many ways.

if you compare Intel's headcount with comparable companies (AMD, Nvidia), you can see Intel is really wasteful

AMD and NVIDIA are fabless. They are not comparable. It takes far more people to R&D a cutting-edge process node and run a dozen fabs 24/7 365.25 than it takes to design cutting edge chips.

If both sides are frosted, then you will have a similar effect as subsurface scattering.

I've been using this for a few weeks now, and it's really handy. But I did learn the hard way that it fails if you don't have internet connection, even if you already have the venv cached.

Not the OP, and not at all an expert in this area, but I was curious what the answer was, and from a bit of reading it seems like a potential reason that might not work is because volatile also prevents caching the value in a register. So if you want to keep the computation in a register, but you want to explicitly clear the value to zero before writing to it, it seems like C semantics are insufficient (at least without inline assembly).

The slowdown is in the simulation, not in the internally perceived flow of time. The internally perceived flow of time would of course not change.

The problem is that an exponential slowdown at each level requires discounting the probability of being in a simulation by an exponential amount per level. So instead of being able to say that the weighted odds of being in a simulation are higher than the odds of not being in a simulation, you have to say that, at best, there's a small chance of being in a simulation.

Photocopying has little to do with simulation of the physical world.

First of all, Bits != Q-bits. You can clone bits. You can't clone Q-bits: https://en.m.wikipedia.org/wiki/No-cloning_theorem

Second, photocopies are static. The physical world is not static.

Exponential slowdowns at each level ruin this hypothesis.

If you think novels, especially good ones, are "stories of made-up people that aren't intended to illustrate any particular idea", then you're missing quite a bit.

This model of physics does make some falsifiable predictions, and there are discussions about how to test them elsewhere.

Unlike string theory, this theory does not have any free variables to adjust. It's either true or it's false.

I, for one, find it to be trivially true. It fits every observation and is the only theory ever posed that doesn't have the "But why those initial conditions?" problem.

I can understand how the "up to an annual contribution limit" part of my comment could be confusing. I added that in an attempt to not be misleading / to circumvent spurious disagreement. I've moved it inside of parenthesis to de-emphasise it.

If you read my comment as a response to the parent, I believe it will make more sense.

---

CU opened the floodgates of anonymous, unlimited political contributions to "independent", but not really independent, organizations which can campaign on behalf of candidates.

The goal was/is to replace our democratic republic with an oligarchy. And we're now in the end game.

False dichotomy. Prior to CU, individuals could contribute to any candidate of their choosing (up to an annual contribution limit). Companies are compased of individuals. CU was twisted logic to reach a preconceived, corrupt outcome.

Verified pre and post conditions would be nice. Especially in the era of LLM's, which can usually fill in the implementation given a sufficiently clear specification.

To clarify what the parent is getting at: Moore's law is an observation about the density (and, really about the cost) of transistors. So it's about the fabrication process, not about the logic design.

Practically speaking, though, maintaining Moore's law would have been economically prohibitive if circuit design and layout had not been automated.

Probably using Go and defaulting to zero unix timestamp.

Except anyone can start a parallel chain, e.g. DogeCoin. The only thing that makes BitCoin special is that it was the first and is still the most well known. It's value is entirely based on network effects and speculation. It's a speculative store of value which can be digitally transfered for a modest fee.

Fiat has the force of the state behind it. This has pros and cons, depending on perspective and use. As a currency, it's superior in many ways. As a store of value, it's in many ways inferior to BitCoin... Until it's not.

No wonder I've seen such a major spike in spam calls / texts.

Because existing LLMs store no more than 2bits of knowledge per parameter, despite having many more bits of precision: https://arxiv.org/abs/2404.05405

It seems that using OOP in Lua can quickly and surprisingly lead developers into a situation where adding new features becomes really hard.

Not just in Lua...

Skimming MISRA C, it seems to be a collection of either extremely limiting or ineffectual directives. An example of each:

Dir 4.12 Dynamic memory allocation shall not be used

Rule 2.2 There shall be no dead code

https://github.com/sakura1083841400/MISRA-C/blob/main/MISRA%...

This provides nothing approaching the guarantees of safe Rust, as far as I can tell.

The first two changes appear theoretically sound, but it's not clear that they would result in an actual performance improvement at scale. Their analysis ignores that a single matrix multiplication is typically used to calculate the Q, K, and V values from the inputs.

The third change looks like it would break causal masking for auto regressive language models. For masked token language models and ViTs, perhaps it's an improvement, though.

4090 should have enough VRAM for 124M param training. Even at float32 precision, with AdamW optimizer, parameters should only be ~2GB (124M params x 4 bytes per param x ~4 for optimizer weight overhead). So there should be plenty of remaining space for activations.