cavisne

The problem is the hardware not the software, and specifically not CUDA. Triton for example writes PTX directly (a level below CUDA). Trying to copy Nvidia hardware exactly means you will always be a generation behind, so they are forced to try and guess what different direction to take that will be useful.

So far those guesses haven't worked out (not surprising as they have no specific ML expertise and are not partnered with any frontier lab), and no amount of papering over with software will help.

That said I'm hopeful the rise of reasoning models can help, no one wants to bet the farm on their untested clusters but buying some chips for inference is much safer.

Glenn greenwald comes to mind.

Google has been around since 1998 with people putting every dark secret and worry into search queries. Not once has someones searches been leaked by employees, despite highly political people working at these companies with a lot to gain by leaking to a journalist or similar.

Likewise with private messages on Facebook, order history on Amazon.

Big Tech has way more to lose than the small "privacy focused" alternatives, and clearly for them to go this long with this many employees its through design not luck.

You'll get him this time! After 9 years of failing.

I think they only claim their "Ling-Lite" 17B model can fit on a single 96GB GPU, their 300B model needs 8 of them (768GB of HBM)

They talk about CUDA level tracing in their framework. I assume its just consumer GPU's that Nvidia say arent meant to be used in datacenters.

The Fire devices seem like they are on life support from a technical point of view (even if Amazon is making money from them). Their fork is really out of date.

Google is the best of these. You either pay per token and there is no training on your inputs, or it’s free/a small monthly fee and there is training.

Later in the article it says they already plan to leave, as the hydro plant will be connected to the broader grid and make more money selling into the grid.

Doesn't seem like the best donation of $3M to build the plant in the first place, barely connecting anyone and most of the energy going to a for-profit bitcoin miner. $3M is a lot of solar panels + diesel generators.

So to do an efficient MM on AMD you need to find every MM in the pytorch model and replace it with a call to this library? Seems like something that should've been fixed years ago.

Also I assume nvidia does the same thing but it is still hilarious that this is how it works

https://github.com/ROCm/aiter/blob/main/aiter/configs/bf16_t...

Yeah this is AMD in a nutshell. A bunch of fluffy descriptions and then the only concrete example would clearly never run.

EDIT: They fixed the code pretty quickly

A dead user doesn’t buy any more fentanyl? Seems very logical

It's funny to read these deranged comments as at the same time as the astronauts are returning to Earth the only (non Russian) way possible... with SpaceX.

Tesla literally built the engine for that Rav4?

As the author mentioned the rise of ebikes is counteracting this, however the ones popular with teens (Surrons and other similar chinese brands) require you to cut one wire to remove the speed limiter and then they can do 50MPH.

The idea of common spaces in American cities largely died during & after Covid. Anything you did in public (gyms, eating out, public transport, apartment common areas) was suddenly prey to whatever idea a politician had that morning. I think we all remember walking into a restaurant with a mask on, to promptly take it off as soon as you sat down. At the same time as these guidelines were being piled on, basically all laws below murder stopped being enforced. So you had to negotiate a bus full of drug addicts shooting up and riding for free, or leave the sidewalk because it had been taken over by a tent city.

Personal transport is required in America, the places with the infrastructure to avoid it dont have the capacity to govern (any big city). The places with the governing capacity (suburban, rural areas) dont have the infrastructure.

[2] Is an incorrect interpretation. There is no nationwide data on murders by illegal immigrants, only some states like Texas collect this. CPB does not deport many murderers because they are serving prison sentences in the US. In 2022 there were 67 Texas murders by illegal immigrants, more than the supposed CPB FY22 number just in one state.

https://www.cato.org/sites/cato.org/files/2024-06/Policy-Ana...

The per capita trend is still directionally the same as your claim in Texas.

The problem with illegal immigrant murders is there is a legal and proportional way to prevent all of them - don't let them in the first place. This isn't so easy for other murderers.

Is there an explainer on how this stuff works on platforms without some sort of DRM hardware? Or does it not work.

Chrome could probably make a huge amount of money by doing what people assume Google does but actually doesn't - selling users browsing history.

Google uses a complex anonymization/privacy framework to collect some aggregate signals from website visits, but they don't use it directly.

Regulators don't understand this, and technologists who do tend to distrust Google anyway and think they might secretly be using it.

There are all sort so other sketchy things, like what Edge does injecting itself into websites so Microsoft collects affiliate revenue.

There are countries where this wouldn't be allowed, but Google is largely self regulating in its biggest market.

All this would lose Chrome some market share but they are starting from a very dominant position, and for the general public it wouldnt be a big deal - people are already convinced that iOS and android devices are listening to them at all times for ad targeting!

Thats nothing during peak ZIRP people would move tech companies before the launch (or after a project was cancelled) and still ladder climb. "Failing upwards"

This is almost certainly itself an AI written article.

Its funny how Gemini consistently beats googles dedicated document API.

It's based on the most expensive source that is required to meet demand. If there is enough wind to meet demand then gas is not needed or factored in to the price.

All markets (not just energy markets) work in this way.

There is a chinese blogpost from 2019 about 3FS so it predates deepseek [1]. It will be interesting to see the benchmarks but I suspect without 3FS smallpond is not that useful (the bottleneck would move to the networked file system).

None of the big US clouds support Infiniband broadly (Azure & Oracle have some support) so 3FS itself is not very useful to US companies who want to use public clouds.

[1] https://www.high-flyer.cn/blog/3fs/

Not even data, html snippets with links to publicly available data.

Watch at the media turns this into “doge bros mapping top secret minerals” and “doge has access to all our DMs!”

For those who hate Trump but know what the you are looking at in the screenshots keep this one in mind. The media hates you and their goal is to sow division

My favorite theory is he was so unhappy with what liberals did to his LOTR series he became a republican.

In reality its probably because his number 1 focus is Blue Origin and he needs government support for this.

Has it ever happened before? Pre DOGE it was a common assumption that they could not be fired due to various protections. Just because there is a theoretical process doesn't mean it can be done (or has ever been done within a presidential term)

Eg. https://www.cnn.com/2024/04/04/politics/rule-protecting-fede...

Non-probationary employees can't be fired (and even voluntary resignations are rate-limited by the Iron Mountain mine shaft elevator). So every day that process takes lowers the number of employees that can even be fired. Once it gets to a year (very easily) they would all have come off probation so nothing could be done.

Everyone saying "oh something should be done but just do it slower" is actually saying do nothing.

So far Claude Code seems very capable, it oneshotted something I couldnt get to work in cursor at all.

However its expensive, 5m of work cost ~$1 which.

FTTH infrastructure is very low maintenance once built, its mostly passive fiber optic cables that don't degrade. The electronics on either end last a long time too.

The issue with FTTH is the build can be very expensive, you need to either work with an existing utility or dig trenches through an entire neighborhood, and each end user install is a mini construction project.

Thats why community FTTH makes a lot of sense, a small team does that negotiation with the utility, and uses that to be build a business case.