kurthr

If I have a public platform and refuse to sponsor/publish/platform your speech, that is my free speech right! You don't have the right to force me to speak your words. Only the government is prohibited (by 1st amendment) from censoring/limiting speech (that does not cause or promote violence or injury to others). There are also some restrictions and protections also offered to the press (eg libel and defamation) and common carriers, along with the abomination of the DMCA.

Personally, I don't consider paid speech to be free speech, and I don't consider corporations to by humans, but the supreme court does.

Remember, facts aren't real, there is no truth, and crime is legal (in the current "Justice" Dept).

The Japanese have been obsessed with it since the study came out and have a variety of treatments (tannins in persimmons and greentea) though none have passed large scale clinical trials :)

I was fed a YT video about it recently so it must be part of the "moment". Also a topic on menopause blogs from searching.

https://asayujapan.com/blogs/asayu-blog/persimmon-soap-answe...

https://miraiclinical.com/pages/why-this-ucla-trained-anti-a...

So following this to the obvious conclusion, the NSA is responsible for the closure of the Straight of America, high tarrifs, dropping employment, and high gas prices?

You modify tens of thousands of tubes and they are installed as they are replaced in the target facility.

I have to say that after reading other commentary about the 2015 paper this was based on, it seems to be far less relevant, even if it comes up very high in tehGoog search rankings.

There is also the problem that that impedance mismatch between the gas in the tube and the glass surrounding it is huge so very little power is coupled. Interesting things would be for the tube length/resonance to be closer to the sound frequency of interest (eg 200-2000Hz) and close to critically damped. I have no idea if that's the case.

The second question is whether one can simply modify the tube/ballast in a visually imperceptible way. For example, one could insert an IC in the ballast that modulates the power draw of the lamp dependent on the ambient coupled acoustic/vibration power (eg possibly using the lamp as a vibrational microphone). Then pulse modulating the power draw of the lamp at a relatively low noise part of the power spectrum and within the band of the main power transformer filters. The lamps easily draw 40-80W so there's lots of room to work, and there are already Power Factor Correction implementations that use this simple technique to make lamp's inductive loads "look resistive".

So what's notable is that either the ballast of the lamp could be tampered with or one of the bulbs. Either one could allow transmission out along power lines, if standard transformers and capacitors are used for power delivery. This is how home powerline ethernet works and it transmits Gb/s rather than the few kb/s needed to get voice out over the power lines.

Case in point. I saw a single Korean Samjung lamp interfere with AM radio stations (<1MHz) from 100 meters, by just clipping the power draw of a small fluorescent with a thyristor and oscillator at the ballast of a desk lamp to provide PFC so that thousands of lamps could be placed in huge open plan offices.

Exactly, what exactly is an English major going to do, if YA novels are the pinnacle of reading achievement? I'm baffled, and reading more about it, concerned.

I've talked to a few high school and young college students in the last few years to get a feeling for what pandemic and LLMs have done to education, but this seems near apocalyptic to me. An inability for a majority of English majors to read 19th century literature, or for their teachers to even bother teaching with books? I think the students I've been talking to have given me a false sense of hope.

https://nataliewexler.substack.com/p/struggles-with-bleak-ho...

the "soft bigotry of low expectations" has come for AP classes and college curriculums.

If you're worried about >100mg (0.1ml) then, don't think about the dosages people use for botox! It's 1x,2x,3x 80picograms per injection.

I guess it's a biological product, but it's certainly not "living".

Then there's carfentanil and similar which are deadly only at much higher 100 microgram levels.

How much does open source rely on copyright for enforcement?

Do court decisions that limit copyright protection of AI generated content cause problems for open source? Isn't there a requirement for a "Mental Conception" of the code output when using such a tool? Even copypasta requires you to somewhat know what you need to do, where "automated bug finding", does not seem to.

From the outside, it seems like Canary Warf was where most of the high density building in London was done. That wasn't nearby lots of low rise single family homes, was it?

Certainly, some people are moving to the Bay Area, but the rate of home building well exceeds the population growth rate of California (or England) which are occasionally negative. It's mostly a question of new people wanting to move to particular places they like. But ask, why do they like them, and how was that created? Also ask, what fraction of the residential homes are occupied. In California (especially in desirable areas), the answer to that is likely <90%. Nobody (especially not developers) want to raise taxes on empty houses/apartment/condos/retail, but maybe it would fill up if they did (and lower prices for people who want to rent/move in.

All of Tokyo was basically rebuilt after the war, because it was bombed. Much of it was not originally built to earthquake standards. Rebuilding high density in bombed US SFH neighborhoods, would certainly make sense.

I completely disagree with broader point of "cultural reasons" why Asia built high density, because it is all so recent. Mostly in the 1970s-2010s. They built entirely NEW infrastructure, with transit (e.g. subway, highspeed rail), malls, streets, commercial and residential buildings. Look anywhere in Korea, look at the rise of Shin-Yokohama (outside Tokyo), look at the highrise Urbs outside of Paris (out all the way to Rennes).

Why is the center of Paris still only few stories high after the Hausmann era when it was laregely built over 150 years ago. Why don't they put 50 story skyscrapers in Paris... because they don't want to, and they push the "density" to the edges where new facilities can support them. But the housing prices in Paris are astronomical (just as they are in London).

If you want to talk about closing schools, because there aren't children for them... then great! Build new high density where the school used to be. Build a school into the building, put a subway under the school with a connecting mall. Then at least the surrounding neighborhood that has been paying 70 years of taxes doesn't have to deal with the extra traffic/parking of the new construction. Maybe they'll even enjoy the new mall, and foot traffic will help support existing commercial infrastructure.

So you're saying that they build giant malls, office building, and (what is left of) manufacturing in cities by tearing down SFH?

Certainly, not what I've seen. So people coming for "the jobs" DESERVE the kind of nearby, low cost, transit friendly housing that does not currently exist, while the people who live there DESERVE to have it taken from them at a price lower than the housing will currently bring?

Why?

But why does the high density housing have to be where the SFH are? Why does it need to share the parking, facilities, retail, roads, schools, etc. of something that was built and has been maintained and evolved for 50-100 years? Why not build the high density with it's own transit, roads, malls, parks, schools etc?

This is how it's done in the rest of the world, certainly Korea, Japan, and China. I see this in England and France as well. They develop a high density area with new facilities that can grow and evolve for how people want to live there, rather than mowing down houses (in communities of color) or plopping 5-7 story apartments without parking or parks right next to single story SFH, as developers like to do in the US.

I assume it's because it's more profitable for the developers. They don't have to build anything, but the part they can immediately sell, and they ride on the value of the neighborhood created by the people who already live there. I think some infill near commercial districts or malls also faces much less "NIMBY" criticism.

To give a more personal example: I like the location and view of where you live. I like the plants you water, and you keep the house looking clean. I want to just build another large room on top of your house, where me and a bunch of friends will live. We'll just share the stairs, hallways, and front door, and of course our share of the water and electric bill. We probably won't be too loud most nights, and usually when friends come over they won't use your bathroom or take anything from your rooms, or at least we'll ask them not to. Sorry about the construction noise while we do it, and just think you'll get to meet so many new people! Sorry that the part of your house you still own will be worth less than your neighbors when we're done, but that's the price you need to pay for living in a nice place and keeping it up.

The argument that SFH infrastructure repair is subsidized by new building is a much better one. Of course that was true of the tax increases and subsidy of the original town, when the SFH were built as well.

I'll note CoreWeave didn't commission the very first Blackwells until ~Jan 2025 so it hasn't been very long for RMAs. Furthermore, for training even when the NVL backplane can use the remaining GPUs a ~20% drop in performance means it isn't in the training cluster.

I've personally heard this from several sources in the data centers (installers, training, network). It's not uncommon for 10% of racks to fail on delivery. I hear that's improved somewhat from GB200 to GB300, but the number of FW updates from the time they ship, until they're commissioned is >>10. If an HBM or GPU or backplane supply/cooling fails, it is basically not swappable or repairable. You have a "dead" rack, and deliveries are on allocation so you don't get a replacement for months (eg some "RMAs" for early delivered parts in late 2025 are still dead racks 9 months later). "Tray" swaps are technically possible, but still quite rare, perhaps because debugging takes as much time as commissioning a new rack.

I don't want to out anyone, but these are similar comments:

https://www.linkedin.com/posts/neelmaster1_aiinfrastructure-...

https://www.hostzealot.com/blog/news/nvidia-gb200-nvl72-is-n...

https://introl.com/blog/gb200-nvl72-deployment-72-gpu-liquid...

Since the current NVL72 are still at the ~15%/yr failure rate it's not clear your new data center is going to have half it's compute in 3 years. If you're still running H100s they draw >10x kWh/Mtoken as new designs. All of these systems become dated, but not all of them require entirely new infrastructure.

If a ROM rack running a near frontier agent model at >10ktoken/sec costs <$1M (rather than $4-8M for NVL72s) and draws only 10-20kW (rather than 100-200kW), and doesn't require a completely new cooling and power system every time you update? There'll be lots of demand for GLM5.3 in a year.

What these don't do is TRAINING, they only do INFERENCE, but they could do it pretty well.

The metal masked ROM is basically only 2 metal/contact layers. It's not a full new design and tapeout. You could roll a new set of parameters every ~2-3months. It's not an architectural change. See statements below.

https://www.eetimes.com/taalas-specializes-to-extremes-for-e...

https://www.turingpost.com/p/taalas

https://cambrian-ai.com/taalas-launches-hardcore-chip-with-i...

Part of the key is that by moving even from 6nm to 3-4nm one could embed a 20-30B model as part of a MoE (or only a subset of activated layers) on a single reticle die (note B300s are already multi-reticle), with a separate predictive/dispatch model controlling them each on a separate chip. This is without even stacking CiM ROM die. Moving the layer activations (and KV cache etc) between die requires relatively high speeds (and low latency), but distributed with multiple die in parallel might well be doable even with standard multilane PCIe. Of course KV cache prefill could also be handled by external GPUs. I'm sure AMD will make some reasonable choices.

Contractors are commonly net30/net60/net90 or paid on completion and inspection of work done. This isn't a union job with biweekly paychecks.

Yet strangely, we imprison people based on their intent all the time.

For years a woman ran around in traffic and was killed by a car to protect her from the state. Nothing could be done? She had to die?

https://www.sfchronicle.com/bayarea/heatherknight/article/Ca...

He didn't didn't want to be the SECOND person squatting with boots!

There is an oral PCSK9 treatment (enlicitide) was recently released that provides a 30% reduction, and of course you can't get insurance coverage without the test. And there are 2 drugs with 90%+ reduction in Lp(a) in phase3 trials with expected release in 2027-8, in particular Olpasiran.

Yeah, (Optane) 3D XPoint from Micron would have a decent density speed trade-off. I don't know whether it makes the margin they can get from RAM though. Or whether CiM for masked ROM with a 3mo wait for parameter update (and much lower power) is a better trade-off for inference.

They must each have 20mm customers paying $10 a month to have 50% margins in profit. Or are they just VC funded?

Are there 100 of them? That would be 2 billion customers.

So thinking this through to come up with your numbers, total investment through 2027 is ~$2T and interest on debt is over 7%, which makes servicing this ~$140B/yr. However, failure rates on H100-B300 installed HW have been over 12%/yr even as the power and cost efficiency per token of the later builds has risen ~5x. So depreciation on the data centers is conservatively $240B/yr ignoring power costs (likely only $30B/yr at $0.05/kWh). Conveniently, this is $20B/month and if 5% of the population uses it, that's only 400million people so it's $50/mo or $600/yr, only if the AI-vendors make zero profit and $100/mo if they make 50% margins. Realistically, the OpenAI and Anthropic go to zero or it's $600/year.

I don't know where the 5% of world population came from, because that's clearly not just professionals or people making a lot of money. That's Uber drivers, and retirees in the developed world or tech workers in Asia making <$10000/year. Those don't look like great markets. This needs to be 2x higher value than their cell phone and internet that they might spend $300/year on today (that's a new iPhone every 3 years on an ATT plan). It's not like it can replace their plan, because they need that connectivity to use it!

Who's getting this value other than SWEs? There aren't 40 million SWEs and I don't see them spending over $6000/year. If their business does, it still has to pass on the cost to consumers and/or fire SWEs.

Yes, but that is the point we don't anthropomorphize stochastic outputs even when they are dangerous.

Obligatory quote:

"Do not fall into the trap of anthropomorphizing Larry Ellison. You need to think of Larry Ellison the way you think of a lawnmower. You don't anthropomorphize your lawnmower, the lawnmower just mows the lawn, you stick your hand in there and it'll chop it off, the end. You don't think 'oh, the lawnmower hates me' -- lawnmower doesn't give a shit about you, lawnmower can't hate you. Don't anthropomorphize the lawnmower. Don't fall into that trap about Oracle." — Brian Cantrill

It does give you the option to change cards when you double tap, before you pay.

But I think the real reason is that people would walk around with card readers and scan peoples watches to collect cash. I can only guess it doesn't happen in Japan for cultural reasons. You could try and do that with purses/wallets too, but since most people are carrying multiple cards, they interfere and largely prevent RFID skimming.

You know the other killer feature of the Apple Watch? Find my phone!

See someone with an Apple Watch on frantically looking for something, just say "Hey you know your watch knows where it is". Just press the button, DING.

I've read a lot of these, but many that I've never considered masterworks. I mean, I'm a Dick head, but that's a lot of representation on the list.

What's really interesting is that if you scale it to higher densities (eg 3nm and stacked die) with ComputeInMemory for fp8 you can reasonably start to fit 30B-70B models. With MoE and multiple stacked die, just like HBM, you could fit an open weight near frontier 1T model (like GLM5.2) at similarly much lower power <10kW and high token rates >2ktps. For running a bunch of agents where fill rate and speed/latency are important it may not matter that you're 6-18mo behind on weights. The process for the chip design could be largely automated, and new silicon pumped out as new weights are available (with a 3-6mo delay).

Good food, good meat, good god, let's eat!