In China street markets you can get a 48gb 4090.

I’m not sure if equipment to make older ram types still exist, or if it’s possible to use ddr3 chips on ddr5 interface/voltage, or gddr5, but this is getting too stupid.

HBM prices are also rising fast, and next gen datacenter AI cards are likely to be worse value (flops/bandwidht/ram) than older ones. But still ddr5 would be more profitable to make than HBM at half the current price. So everything is stupid extortion, and should come back to some parity between the 2.

If you ever stream cable news in the background during business hours. experts will come on to tell you that if the US doesn’t bankrupt itself, then Russia or China wins. The only remaining hope for US economy seems to be to beat China at AI, because it can’t beat it at anything else. So, all the money for military, and AI to help military make Skynet, which will mostly be used to ensure establishment can control/kill Americans who don’t support Skynet for Israel supremacist rule over the US, is simply a continuation of the tautology of “all US money for militarist US establishment control over all of you”, but with more efficient AI.

But, but… surely, the most important AI application of all time, my OpenAI girlfriend, who truly loves me beyond all monetization opportunities, will never be replaced by a hotter newer model.

UBI is not going to win out over soylent green.

Establishment prefers solylent green solution to unneeded slavery for sure. UBI is an obvious winning political platform, even though it does not provide the proponents with maximum political power, the way that bandaids on oligarchist supremacism still allows for supremacism and soylent green to rise as the solution.

Simply, the answer, is to reject establishment politics, including Bernie as controlled opposition proposing nonsense that gets shot down for its stupidity, instead of UBI.

Also, an MIT study released today, said 135m US jobs could be replaced with AI agents.

The only possible customer for committed datacenter projects is the US government. the bailout is just buying all the datacenter time… because if not, China wins. Amazon, despite no congressional appropriations yet, has announced $50B datacenter to supply government customers.

Not sure of exact windows comparisons, but 5090 had problems, and I don’t think amd is ahead of nvidia by as much on windows.

If we don’t bankrupt the entire economy to achieve datacenter Skynet, then China will win. These prices are reversible, and you should boycott for now instead of hoping to resell to a bigger sucker later.

NVIDIA high inventory levels are most likely responsible for this. They keep ordering more HBM. At same time, they aren’t selling that much more, and giving big accounts receivable delays to their concentrated big customers. Hourly H100 rental prices are down 33% since august, and so even without price hikes on datacenter cards (which is certain due to massive ram, and ram prices that have to affect AI cards too) actual demand may be going down which is what will shift memory production to more useful categories.

The case for total US collapse, is government policy that steps in to buy all of the datacenter time for Skynet goals. Already, make NVIDIA richer by increasing their sales to Zionist allies and China is to make every American poorer. The unanimity of evil in US political/media establishment stays quiet on the heads, destruction of America through bankruptcy; tails, Skynet slavery and genocide.

There’s a scam by large companies, Apple included, where pricing globally is based on US prices. Even with tariffs by US on everyone, and no tarriffs in US colonies, companies boost prices everywhere to subsidize US customers. Our colonial governments say nothing.

has to pay a per-device licensing fee.

Where I’m confused, is that it would be a perpertual/long term annual license fee per device. It would make sense to have a one time fee per device shipped. That would not affect older models.

I guess what is happening is that manufacturers can stop paying for the capabilities by “downgrading” their driver support, and it affects old and new systems the same when users “update”?

does dell/hp have to pay annual license fees in perpetuity for systems they sell???

if tsmc only makes datacenter chips from now on, then “we” are shut out from the huge privacy (and fine tuning specialization) gains given by small efficient cheap to run models (or play games on new hardware). US datacenters will serve US empire/political establishment both with government as main LLM customer, but also for data collection/palantir ontology/social credit scores on every American.

I suspect that better datacenter chips won’t actually reduce their cost due to supply limitations, but even for small efficient models, personal hardware has a long payback period compared to a per token “rental” cloud charge. It is unlikely that all of the datacenter chip buyers will have non-government customers to use them all, and so either bailout or bankruptcy followed by megatech buying the datacenters for cheap followed by a bailout in government revenue for big tech global/citizen control applications.

Eventually, even the government has too much AI resources, at planned expansion pace, and then consumer/business computing/gpu market comes back. Could be as soon as 2026 that a collective understanding of absurdity occurs.

Chinese emerging players who are shut out from tsmc/taiwan/ROK chips and memory sources.

This is extremely serious for economic bubble.

Orders for datacenter AI chips exceed supply, and more high end/other memory per TSMC wafer is further nightmare. This is likely to mean higher prices per token for datacenter buyers, and higher prices for users/model renters, and much slower demand growth and AI progress. It also means long delays for datacenters, and better black market (China, higher than MSRP diversions from contracted deliveries).

I’m not sure if affects phone/lpddr soldered memory, but tsmc is going to charge more for phone chips too. This can cause whole consumer/business computing market to collapse. Return of older generation designs on underused process nodes will give little reason to upgrade, and still overcharge. This can be an opening for China exports of competing products that were not possible at low/reasonable ram/tsmc prices/availability, where even if China has difficulty achieving best yields, it’s still profitable to invest/expand aggressively, that discourages US/western colonies from investing.

This race to give the US Skynet, for stronger political control/social credit/surveillance of Americans, can make a bubble in everything else, and accelerate financial collapse, all the while making the goal impossible to achieve and forcing China to become stronger/more resilient, with greater share of global computing supply.

seems more mature, and external automation oriented. I’m interested in developing voice text editing tools. Don’t really know if what I linked is best suited platform.

System already demands recipient updates every 6 to 12 months.

Another good argument for UBI to neither burden an administrative expense nor people on justifying/regulating hunger relief.

30 gt of co2 is the new aspirational net 0 :(. That is still 3ppm per year increase. Feedbacks likely to make it worse.

A classification problem is correlation vs causation. Sunspots and mini skirts have been correlated with stock market returns to some degree, but it tends to be a tenuous connection not guaranteed to hold up over time, or to actually have any meaningful relevance whatsoever. It’s easy to oversell models.

Dystopian neutrality in article.

without discriminating on grounds of protected characteristics

AI classification is trained based on supervised (the right answers are predetermined) learning. MechaHitler for the fascist nationalism’s sake, will rate Obama’s face as a poor employee, and Trump’s as the bestest employee.

Open training data sets would be subject to 1. zero competitive advantage to a model, 2. massive complaints about any specific training data.

For some jobs, psycopaths AND loyalty are desirable traits, even though they can be opposite traits. Honesty, integrity, intelligence can be desirable traits, or obstacles to desperate loyalty. My point is that if there are many traits determined by faces, much more training data is needed to detect them, and then human hiring decision based on 10 or 30 traits matching to a (impossibly unique) position, where there direct manager only cares about loyalty, without being too talented, but higher level managers might prefer a candidate with potential to replace their direct manager, but all of them care about race or pregnancy risk, and then post training based on some “illegal characteristics”.

A Gattaca situation where, everyone has easy time getting great job, and moving to a greater job, OR being shut out of all jobs, creates a self contradicting prediction on “loyalty/desperation” controlability traits. If job duties are changed to include blow job services, then surely those agreeable make a better employee, despite any facial ticks responding to suggestion.

Human silent “illegal discrimination” is not eliminated/changed, but the new capability, you can use a computer to do the interviewing, and waste more interviewees’ time at no human cost to employer is why this will be a success. A warehousing company recently looked at facial expressions to determine attention to safety, and this leads to “The AI punishments to your life will continue until you smile more.” Elysium’s automated parole incident interviews is a good overview of the dystopia.

relative zero is an even number! There I said it, and I don’t care what any of you think of me.

Softbank has a history of knowingingly inflating private market prices with additional small investments at a higher valuation (with syndicated other people’s money) to make their overall investment value appear to be going up. They sell out at IPO or later rounds.