halfdeadcat

After thrashing on it for a day, I could not get it going on DGX Spark with FP4 quantization. I find this irksome, since Nvidia created the NVFP4 specifically for Blackwell. The Nvidia cookbooks for this model are all for H100. I tried vLLM, Ollama and various patches. As of right now on DGX: you can feasibly do FP4 on dense models. But FP4 + MoE is a largely a dead-end.

Good luck getting it to run with NVFP4 on a DGX Spark, the very architecture Nvidia created that format for.

That is incorrect. The article is referring to the cost of the infrastructure, not the cost of the electricity that infrastructure produces. Power plants are spec'ed by megawatt (peak).

It's a 'dragon hatchling' because it is 'scale-free'.

SMES systems offer grid-level energy storage at about 98% efficiency. You could also move nuclear reactors to the middle of the desert because no one wants them in their backyard, then losslessly transport their power back to the grid. Same for bringing power from solar and wind installations. These are just two examples with immediate commercial application. The economics of energy production shift when you eliminate 15% of the cost and distance limitations. That would have a big impact on decarbonization. Longer term, josephson junctions replace transistors in processors.

The premise of this blog was this: the author budgeted 6-8 weeks to tackle a problem that he finished in 3 days. That doesn't make him 10x as productive as someone else. It just means he is really bad at estimating how long it will take to solve a problem. Estimating how long it will take to figure something out is a devilishly tricky problem.

The things that made me feel older than anything was: When they recorded the audio clip for their webpage, they used Massive Attack's "Unfinished Sympathy" as the background audio. I still love that tune. Of course, they couldn't get away with using that now without proper licensing.

This is just a PR release. I hope their claims prove out. But there is no evidence supplied and breathless PR releases seldom go anywhere.

Perhaps if a founder is too delicate to be asked questions grounded in reality, they should consider a different profession.

Not wasting. They are obviously unused blocks allocated to them.

Never hold much faith in any proposed new GA aircraft this early in the development process. This announcement basically is a group of people saying "Wouldn't it be cool if we..."

Interesting article.This virus apparently developed a trick where it uses host proteases to clean its own viral envelope, allowing better fusion with host cells.

Oh good. This will be handy if I ever need to compute FFT's using only Magic cards.

This is a risk Amazon assumes, by offering products that customers cannot see in-person before purchasing. There is no substitute for actually handling the product, and Amazon can't exist if people are reluctant to purchase items for this reason. While returns may cost Amazon, they are essential to convincing customers it's ok to by unseen merchandise. Think Zappos (also an Amazon company).

If they only had some sort of self-contained atmosphere suits.

It's a little surprising because I'm tempted to classify this one as 'requires 0 lines of code'.

So, an exploit using off-the-shelf software. Just awesome.

Maybe it's a hoax. But the pieces seem to be falling according to the plan. http://money.cnn.com/2014/02/24/technology/security/mtgox-bi...

Cup-a-bits

I work 90% remotely. When I started this job, I came into the office everyday, a 40 minute commute. After about a month, my boss came up to me and said "Are you meeting someone here today? If not, what are you doing here?"

I actually like to go into the office from time to time, but there isn't any point when no one is there. I spend a lot of time at various clients' offices, and I notice a correlation: the more opposed to remote work a company is, the more meeting-driven it is. It is as if management feels like it needs to do something with all these bodies taking up space in the office. I see small to mid-size companies sucking up productivity with endless meetings.

That said, I think my company has lost a little bit of interactivity with the remote work force. I'm putting together a Google+ Hangout I call "Engineers' Office Hours". The idea is that if you aren't doing something billable, you hop in and offer help to someone stuck on a problem.

Is it just me, or does PaddleOver, one of their two demo apps, doesn't actually work?