cavisne

“They usually advertise themselves by a user agent string” my understanding is this is very much not true, but all that info comes from cloudflare which obviously has an incentive.

Are there stats on this?

This is exactly the sort of attack this change is helpful for, as pip would not work at all on these machines. Probably the only way to stop these “social engineering” (as google characterized it) attacks.

Pretty sure 1Password has the same constraint. Lastpass did not but well... they dont have the best reputation anymore.

Isn’t it impossible to know if you haven’t had Covid once vaccinated? (Because you will test positive for antibodies regardless)

Imo Google (basically by accident) owns the best place social network for the next decade - YouTube.

Human generated content will become rare and more valuable, which is why the remaining non walled gardens (Reddit and Twitter) are quickly raising the drawbridge.

Video will be the last medium to fall to AI generated content pretending to be human generated, and by cloning tiktok across to YouTube shorts they remove the barriers to entry for humans.

Social media nowdays is not about friends, it’s about algorithms & creator monetization. YouTube is #1 on monetization by a huge margin.

Any text based social media network will be absolutely overrun with LLMs.

I have heard it only applies to Google and Facebook at this point in time. I doubt any other companies would be even capable of complying.

The "even for inference" thing has turned into a bit of a trap imo.

Data parallel models scaled up for training and then could run on individual chips, but these massive model parallel models require a couple of chips directly linked together even to do inference.

So the idea that a competitor could come in with a simple, cheap inference chip doesn't really work.

One interesting thing is if the model can’t fit into GPU memory (sharded across multiple chips) it would be much slower. So one chip would take more than a month.

That’s why these clusters are so valuable, even with Nvidias margins they are still cheaper than using less compute for longer.

Wonder what the outcome for early employees looks like here?

This is not a very big company and hasn't been around for very long. 1.3B is a lot of money

I feel like it must be a bad business in general. The fact that Azure also has a weird setup with godaddy managing the domains.

elastic inference comes to mind.

One subtle benefit of cutting edge routers, a lot of routers nowadays are WiFI mesh routers. This traditionally was not very good, but at least at the moment a mesh between routers over WiFi 6E (which most devices do not support) works really well.

Claude is a mystery/surprise to me. My mental model has been to train these cutting edge closed source models you need 1) Bespoke supercomputer (no public cloud will cut it) 2) Great dataset (which takes a long time to collect unless you have a partnership with with a search engine) 3) Couple hundred lines of pytorch code to run on the supercomputer 4) A couple of employees with experience in the dark arts of malfunctioning GPU's and exploding gradients

Anthropic is a relatively new startup that probably has 3) & 4) from their history at OpenAI. But I don't see how they could have 1) & 2).

The (sad) reality of GPU capacity in public clouds at the moment is your basically need a sales call and a reservation to get any capacity.

So the on demand price is not really relevant, maybe thats why they excluded it.

Limited to 40 miles an hour and only allowed on freeways seem like incompatible constraints...

Its impossible to profitably mine crypto on a public cloud so anyone doing it is either stupid, or has no intention of paying (much more common!).

Historically GCP/GAE has had a more generous free tier without credit card requirements, so they've always been a bit stricter on this than the other clouds.

Not literally though, it’s run by some company that is not Amazon.

I feel like that’s exactly what will happen, and this company would be well placed to benefit.

Basically some European company licenses the “AWS” stack and follows instructions to build and operate the data center.

Works in China, and even more so with the various government regions.

Europe is basically building the Great Firewall slowly and through committee, they’ll end up in the same spot.

The plug and charge protocol is open

https://en.wikipedia.org/wiki/ISO_15118

While what you say is true about a gas station, I think a lot of the standard privacy arguments dont really hold for a car.

Its literally illegal to drive a car without it being built to the standard the government likes, with an identifiable ID printed on the outside of it, with a valid license with your name and date of birth on, and current registration.

Let alone the fact that the vast majority of cars are sold with gps trackers bolted on if they are financed. Requiring a credit card tied to that vehicle is the least of your privacy worries.

Ford has already announced they will support plug and charge on Tesla chargers also

The whole paying on a terminal/using an app thing is the wrong approach I think. Ultimately to charge there is a handshake protocol between two computers, just sort out payment then (and thats the standard NAC supports https://en.wikipedia.org/wiki/ISO_15118)

Adding a payment terminal is just adding another handshake and computer into the mix, a computer that needs to be (mostly) connected to the internet, exposed to the elements, and vulnerable to skimming devices.

Maybe because they have actually used an electric car?

CCS is not fit for purpose, even more so the CCS chargers.

https://www.youtube.com/watch?v=fq0RAjJ1PKQ

They are absolute garbage hardware with a clunky, fragile connector.

And the only company deploying them at scale is doing so because they (VW) were forced to when settling Dieselgate.

The situation is even worse as the cable length on the supercharger and the port placement on non Tesla evs means every non Tesla blocks 2 chargers.

Hopefully something ford sorts out when they embed the port in 2025.

Yes and to be eligible for that they needed to support at least 2 automakers. This gets them there without having to retrofit stations with ccs

Gaming class GPU's provide much better $/FLOPS (ie a 4090 is ~5x better value than a H100)

https://timdettmers.com/2023/01/30/which-gpu-for-deep-learni...

Nvidia forces corporations to buy datacenter GPU's for CUDA licensing reasons, and adds a significant margin on top.

The Uber example seems like a stretch. If 28% of america uses it monthly, its probably comfortably a majority when you stretch to yearly (ie airport trips). Especially if you extend it to include Lyft.

Rideshare has comfortably replaced taxis, likewise Doordash + competitors have comfortably replaced the alternative (which for most of the country was pizza delivery).

A lot of this was manufactured, it came out recently that NYT has a top down editorial policy that tech cannot be portrayed in a positive light

https://twitter.com/KelseyTuoc/status/1588231892792328192

I'm surprised this wasn't a bigger story at the time. News orgs will obviously have a bias based on who they hire. But historically they always try and appear to be balanced at the editorial level, ie Fox News has a few token left wing people, CNN has/had a few conservative punching bags.

Interestingly the actual judgement is significantly higher (still single digit millions). I dont really know what a "Statement of Financial Condition" is but I assume it means they are taking basically all the money these companies + their directors have.

These are not big operations, they were charging 7c a "lead", so each company made ~100k, and the CEO was running the script to send fake data himself.

TPU's have a TPU-TPU interconnect that is faster and lower latency than any GPU cluster [1]. That said this is a huge leap for GPU's on GCP. For A100's SOTA is 1.6tbit per host over Infiniband (which azure and some smaller gpu clouds provided), AWS had 400-800 Gbit and GCP had .... ~100gbit.

SOTA seems to be 3.2Tbit for H100 clusters so this still seems a bit slow? (Tricky as they don't give us a clear number just 10x). H100's are much more powerful per chip though so at least initially the clusters will be smaller and not network bound.

The tricky thing is no one other than Azure of the big providers seems willing to pay Nvidia's margins for RDMA switches, it seems this is still the case.

[1] https://arxiv.org/pdf/2304.01433.pdf

The author makes some good points but I would be wary of the motivations.

Open AI's "moat" is they have got ~400 researchers to work in roughly the same direction, not working on their own projects with the sole aim of publishing a paper. The outcome is an amazing product.

Letting everyone loose with their own LoRa finetuned model that can beat a single benchmark (and make for a great paper!) is probably the wrong move. I'm yet to see any open source model that is even close to GPT 3 (let alone GPT 4) in actual real world use.

Is MLIR a “huge success?”

Is it involved at all in the Python ->PyTorch -> nvidia gpu path that dominates ml today?