And the sad part is that ai would be a perfect fit for intermittent energy sources: run inference 24/7 (on whatever you can muster, even if it's fossil it simply does not consume all that much compared to training) and over-build training to achieve whatever total throughout you need in time of energy abundance. How heavy would CO2 pricing have to be to make the market do this instead of the exact opposite?
What would be the economy of these data centers if they would run using power source with say, for example 25% capacity factor? What is the capital expenditure of a data center? What is the yearly operational expenditure of a data center?
I found that a 100 MW datacenter can cost roughly $3.35 billion, with a significant portion going to high-end GPUs like the B100 or H100. For electricity, 100 MW data center can incur annual power costs ranging from $41 million to over $131 million, depending on regional energy prices.
Compute has massive initial capital requirements, and a very short obsolescence window. It does not make sense to invest in compute and then not run it 24/7 at 100%.
Hardware sitting idle waiting for sunshine/wind is leaving money on the table. Especially if you can burn gas during that time and still be hugely profitable.
Would the world be inherently different if hardware cost was 4x of what it is per unit of training done? The world would be inherently different if we did not burn through fossil stockpiles at the speed we do, or not at all. It's market mechanisms failing to lead to a desirable outcome. Nothing new about that, plenty of problems of that kind have been solved before, but yes, this one is more difficult.
The trade-off between having the field-swappability feature and going the lean way (it's not just cheaper, also smaller, lighter, less to go wrong) shifts though: when regulation forces companies to go 20% of the way towards field swappability, more will take the bet that there might be a niche worth serving at the 100% mark.
I still would not expect this to happen for mainstream phones, but other devices? There will more field swappability with the regulation that enforces layman replaceability then without.
Is it just quantization or is it also rearranging the weights to get clusters with (almost) the same factors? If it's the latter it would very much be training in full precision (but also hardly any precision lost by the compression).
Unfortunately my mental model doesn't contain anything to even guess if that's possible or not, my AI times were at the falling flank of symbolic. Funny how one bit models feel a bit like approaching an approximation of symbolic again (until you read about the grouped scale factors and then the illusion is gone)
One thought that suggests rearranging is not involved,a thought that does not require any knowledge at all: if it did involve rearranging, someone would certainly have added some order by scale factor tricks with linear interpolation by address offset to lose even less precision.
It's not just quantization. I verified that if you naïvely quantize to 1 bit from the original Qwen model (and set grouped scale factors based on what the original model's weights were like), it just spits out gibberish.
> One thought that suggests rearranging is not involved,a thought that does not require any knowledge at all: if it did involve rearranging, someone would certainly have added some order by scale factor tricks with linear interpolation by address offset to lose even less precision.
Status LED and button placement, those would be the the major benefits over conventional boards in non-exotic fields like putting circuits on plants. Would be fantastically valuable for the one-off tinkerer. But yeah, substituting home etching or milling and reflow would also be quite a dream come true.
I guess they consider that a solved problem: when you can drop arbitrary connections without meaningful heating of stuff outside of the connection, just glue your SMD parts wherever you consider convenient and fuse lines to its connection pads.
But practical application would likely stick to more or less conventional boards (tiny ones for sure) and use those ink lines only for where it's needed. Unless perhaps there's an application where crossing over with simple fused layer printing allows something revolutionary from going 3D? But 2D boards are really, really cheap and multiple layers are already giving ever conceivable advantage 3D could give, outside of stuff like antenna geometries.
For one-off and prototyping, an integrated fused layer + pick&place + circuit fuser machine could be super attractive of course: basically bridging the gap between breadboard and production quality. But I really doubt that this device would be anywhere near hobby workshop tinkering range...
Only when the prices raise to the point that low demand leads to actual flight cancellations. The demand for fuel is much less flexible than the demand for tickets.
Bootstrapping will be near-impossible (or incredibly costly) unless they offer inference consumers models with established demand arriving at some least-cost router service where they can undercut the competition (if they actually can). And then dogfood the opportunistic provider side on their own Macs, but with a preference to putting third parties first in the queue. Everything else is just wishful thinking.
Invasions have historically destroyed elites that haven't or can't flee.
Which means business/scientific/... elites that see things coming far enough out are fine, or get out with a large loss. And yes, I'm sure there's the occasional one that's really smart that gets out with a small profit, but I'm sure a large loss is the more normal experience.
Political elites, which is the large majority: people who are rich because they have a role in government, directly or indirectly are fucked.
In stock markets, insider trading is a big no not because it ruins someone's gambling habit, but because the entire concept of the corporation requires a certain amount of trust of financiers in financees. That whole pooling of capital thing, to do stuff that has too high a capital requirement to start individually. When shares are publicly traded, that trust is impossible when holders have to assume that they will be gamed by employee-owners and that would mean nobody outside the circle of those in the know would ever put in money and then you could just give up and declare that publicly traded corporations simply can't exist. "Don't bother investing, they will strip you".
Prediction markets don't have any "natural" reason like that for excluding insider trading. It's just "game designers" crying their hearts out when someone ruins their game by having an advantage.
The employee could not be an insider if his employer did not exists because of a lack of rules against him trading. The prediction market not existing would not make the insider any less of an insider (we are not tking about people inside the prediction market maker!)
> Prediction markets don't have any "natural" reason like that for excluding insider trading
Corporate employees abusing trust are doing it equivalently whether they trade securities or place bets. Government employees, similarly, don’t personally own the country’s data.
In a minority of cases, the information is one’s own, e.g. bets on how many times a person says a word. But most of the time, there is a breach of trust.
> Corporate employees abusing trust are doing it equivalently whether they trade securities or place bets.
But the crime the article is talking about isn't "abusing trust", it's insider trading. Insider trading is defined in a specific way. For one, as far as I understand it, it must involve securities.
It's not just data, right? Power can be abused as well. That person has power to control the narrative and can make a large bet on the number of times he can say the word.
So now it's public servants military power, congressional power, and they look to enrich themselves with making (or lobbying for) decisions which affect the outcome of a bet.
You could imagine an army general that lobbies for the bombing of Iran knowing the president has his ear, and then bets on the bombing of Iran by March 2026.
>Prediction markets don't have any "natural" reason like that for excluding insider trading. It's just "game designers" crying their hearts out when someone ruins their game by having an advantage.
I can think of a few very good reasons you would want to prohibit insider trading on prediction markets. Betting on war outcomes; being incentivised to commit war crimes or throw vital operational goals for financial gain. Wagering on public figures' jobs; being incentivised to harm them.
CFTC's guidelines around prediction markets specifically call out war outcomes.
"As a general matter, DCMs are reminded that section 5c(c)(5)(C) of the CEA provides that the Commission may determine that an event contract is contrary to the public interest if the contract involves, among other things, assassination, war, or terrorism."
> Prediction markets don't have any "natural" reason like that for excluding insider trading.
No, they have a different reason. Consider the consequences of a sufficiently large prediction market bet on whether or not <head of state> will be assassinated this year.
Consider also the consequences of insiders and decisionmakers having an immediate financial incentive to take the other side of a bet whose outcome they control.
Of course there's a natural reason: conflict of interest. If someone bets on a release date or a product being cancelled, then they can gain for reasons contrary to the good of their employer.
Corporations depend on controlling the compensation to their employees in order to incentivize them to produce benefit to the corporation. If there is an uncontrolled route to compensation via a prediction market, then the corporation loses its ability to trade compensation for alignment with its objectives.
> The prediction market not existing would not make the insider any less of an insider
Correct, but the prediction market existing makes the insider less of an employee. (Actually, the prediction market not existing would make the insider more of an insider, in that if insider benefit via prediction markets is unregulated, corporations will be forced to limit the information and authority of its employees.)
Chances are they were expecting the agent to spoon-feed hundreds of influencers.
reply