shikon7

If they have 10x cheaper compute, and a lot of cash to spare, why isn't Google dominating AI, like SpaceX is dominating the launch of satellites?

Does Google not want to, or is there another reason they cannot?

It's slower than using just GPU RAM, but it's still faster than using a regular PC that has a much more limited bandwith between the main memory and the GPU. It's some sort of middle ground on how fast and how cheap you can do inference with LLMs that don't fit into a consumer GPU's RAM.

No, it is to make GPUs more expensive for users with professional use cases in mind.

That's exactly the point. The limiting factor to space travel is how to achieve acceleration, and not the acceleration limits of the human body.

That's why he works so hard to colonize other planets.

You need enough RAM to store the model and the KV-cache depending on context size. Assuming the model has a trillion parameters (there are only rumours how many there actually are) and uses 8 bit per parameter, 16 H100 might be sufficient.

On that premise, knowing too much might literally starve you.

I would think that scale would actually greatly benefit journalism. Why is there no Netflix or Spotify for journalism? I'm thinking of something that provides affordable, comprehensive, high quality and individually recommended articles, at a scale that provides large sums of money for journalism.

A single newspaper subscription is often more expensive than Netflix, offering far less production value. No wonder classical journalism is losing market share among young people to more affordable and relatable formats.

In a working market, we won't run out of power, but power becomes so expensive that it's no longer viable to use most of the GPUs. The boom shifts to power generation instead, and we will have a similar article "The power investment boom", where people will debate that we will stop building power plants because we've simply run out of GPUs to use that power.

This was introduced a few years ago in the Swiss retailer chain Migros. The span between the sell-by date and the use-by date (typically around 1/10-1/5 of the entire lifespan of the product) would give the consumer the safety that the product was usable for a few days without checking the dates.

Without it, you need to check the date of every perishable product you buy if you want to consume it even the next day. If the store forgot to remove the product for a day (which happens regularly), you would buy a potentially harmful product if you don't always check the date.

On the positive side, the food can be sold regularly just until its expiry date, preventing food waste.

From the documentation of the differences with Python:

> Strings: Codon currently uses ASCII strings unlike Python's unicode strings.

That seems really odd to me. Who would use a framework nowadays that doesn't support unicode?

Patents usually run out after 20 years, much faster than copyright, and Monopoly was around for much longer.

Considering a ship can carry tens of thousands as much cargo as a car, the comparison isn't even remotely fair.

I wonder if at some point they regret restricting Unicode to 1,114,112 code points.

Why would you use hydrogen to extract the energy of the iron? Wouldn't it be more efficient to burn the iron directly? Likewise, is there no better way to reduce iron oxide to iron than by creating hydrogen first?

Maybe you need it if pointers are 16 bytes?

As stated in the GPL: The “source code” for a work means the preferred form of the work for making modifications to it.

Arguably having the weights and code to execute the model available is the preferred form to modify a LLM.

It’s interesting that night light exposure is so important, but the much brighter day light seems to have almost no effect.

Using compiler specific extenions isn’t really C, though. It isn’t portable and there is no specificed ABI. Add enough of them, and it might become a new language “C with Memory Safety”

You could try to use plants and a growing lamp for them to do photosynthesis.

Yes, it seems the more efficient use we have for compute, the more valuable compute is to us, and the more compute we will be able to afford.

> Encourage fairness and kindness, and discourage hate

> Don't try to change anyone's mind

That seems inherently contradictory to me...

If you ban Google from the internet, many people would still use Google, but stop using the internet

During the night, there will be 1 lane open in each direction (one on the side of the bridge, and one on the opposite carriageway), so the bridge can be moved.

Because Chinese law forbids to convert their coordinates to WGS84, and using the Chinese coordinates for satellite data would give a discontinuity at the border. This is explained in the article.

It seems straightforward that finetuning an LLM just for usefulness will reduce alignment (and the other way around).

Still it is mentioned in the ABI specification, although as an optional type.

It says 34% of adults lacking literacy proficiency were born outside the US, but not how many born outside the US are lacking literacy proficiency, and neither how many born inside the US are.

It seems a bit misleading to state two similar facts (total rate vs rate of outside born) in such an incomparable way.

I disagree. Surely Redis will take advantage of a solution to a given problem.

Rather than assessing whether the LLM has biases, the leaderboard seems to assess whether the LLM affirms the tester’s biases.

Not that I blame them, as it’s probably impossible to define what a “no bias” exactly means.