My experience thus far with the various LLM chatbots is that they've all been trained with the default idea that users want to one-shot solutions.

A human in a similar situation normally would ask questions to clarify the problem, and attempt to ascertain the level of domain knowledge so that they can intelligently communicate.

Part of the song and dance in forming a prompt is to pre-load the model with all those items, so that you don't burn an obscene amount of tokens for it to finally get to the point where it can be useful. An open ended query can be useful for discovery but only if you're willing to do the legwork to cross-reference against other models and reference material.

To put it another way... most humans are lazy and just want to put in the minimum of work, and are willing to accept a 60% acceptable answer from a bot (basically using it as a search engine replacement). Add the fact that most humans are not domain experts, nor are they able or willing to educate themselves to a minimum standard, they have no way of judging whether the bot answer is even at the 60% standard - instead they do what most people do... if it sounds good, and the bot claims to be an "expert", then they go with that.

I anticipate that models specifically curated for specific domains will become more and more useful. I've seen models plugged into residential and commercial electrical code, for example, to act as an always available code compliance helper for a specific municipality. And, as you point out, as long as you can look up the relevant section of code to verify that the LLM supplied advice is correct, and do the load calculations to make sure the math is correct, you can validate the LLM's answer.

The demand is real... whether they can scale the workload down to actually make money remains to be seen.

Long term there are major benefits to society if they can drive down the cost of energy, cooling, and computational infrastructure to support democratized access to LLMs and diffusion models without bankrupting providers.

In the meantime, I would just settle for a decent search engine that didn't require me to go to page 3 of search results (and forcing me to verify that I'm a human for daring to click to page 2 of the search) to find a relevant link that wasn't SEO spam. You'd think that a university agricultural reference for the public would be in the top rankings when searching for a specific variety of vegetable, but apparently today, that's no longer the case.

Given that many companies do a terrible job of maintaining and refreshing their code base... and that even their production code bases are in perpetual beta... and even before AI they were doing things like discarding their legacy codebase and attempting to rewrite things from scratch (remember Sonos?)... an argument can be made that if you're going to churn out a shitty replacement, but can do it faster, that will still save you money.

https://arstechnica.com/gadget...

"In May, Sonos updated its mobile appâ"to the dismay of many users. With missing features and bugs, customers complained about a loss of functionality and hardware not working the way it should. As Sonos deals with the expensive repercussions, a report from Bloomberg today highlights how Sonos allowed the release of an update so buggy and incomplete as to overturn its goodwill with long-standing customers.

Illustrating how poorly this app update has gone, last month, Sonos CEO Patrick Spence said the company would spend $20 million to $30 million in the short term to get the app where it needs to be (which is, basically, functioning as well as the predecessor) and rebuild customer and partner trust. Sonos also expects to miss its annual revenue target by $200 million. This is partially due to its delay of two hardware releases to focus on the app. Bloomberg noted that âoeSonos shares are down 25% this year.â Annual bonuses and merit-based raises have also reportedly been canceled."

I have no idea how much money Sonos spent before committing to spending an additional 30M, but taking a 200M revenue hit because they couldn't get enough feature parity before shutting off the old servers and disabling the old app seems like a pretty big stick.

"One reason for the appâ(TM)s failure is the outdated code and infrastructure that the prior app was running on. Anonymous employees Bloomberg spoke with claimed that the Sonos appâ(TM)s technical debt had been building up for 20 years before the update.

By the time Sonos decided to update the app in mid-2022, it was dealing with software based on virtually obsolete infrastructure and code languages. As such, the app update âoewas less about introducing new functionality than sorting out the existing mess,â Bloomberg reported.

After decades of the appâ(TM)s inner workings growing stale, the impending release of Sonosâ(TM) long anticipated Ace wireless headphones, which came out in June, made the need for a new app both urgent and necessary. This is because the headphones were made to be on-the-go, differing from Sonosâ(TM) other products, mainly speakers and soundbars relying on home Wi-Fi. This seems to align with comments that Spence made to investors in August. He said that the app update was âoea redesign of the entire systemâ"not only the app but also the player side of our system, as well as our cloud infrastructureâ"and this was a complex undertaking.â"

This sounds like they didn't want to support parallel systems and were hoping that they could migrate everything over before the cutover... and when they failed to make that deadline, they decided (and by they, I mean management, because who else in their right minds would disable functionality to happy customers) to just deep-six the legacy environment without warning instead of transitioning people to a "legacy retention program" with a definite cut-off date in the future.

I guess the moral of the story is, if you have shitty management, having them make fucked up decisions that result in near-immediate feedback is better than having them percolate for a year and then result in no-bonuses / layoffs.

You don't own it.

You can extend this to hardware as well. If you're not allowed to repair or modify something... you don't own it.

This reminds me of microtransactions.

We're 50 years overdue for an alternate payment clearing system that doesn't take more to process than the payment is worth. Most of the enshittification we've seen due to an eyeball (now attention) based economy could have been sidestepped if we built a system for people to pay on a transactional basis for compute and content.

Funny how the wave of autonomous agents is once again driving people to revisit this stuff.

https://spellboundproductions....

"Thirty years ago, General Magicâ(TM)s Telescript established a foundation for autonomous agents. The vision behind it sounds like today's promises of Agentic AIââ but in fact was actually enabled by the technology from the outset. Telescript didn't ask users to just trust it. Instead, it built deterministic safety into the system itselfâ"Permits that limited resource consumption, Authorities that enforced accountability, an Engine that prevented agents from touching host resources directly."

There's no reason you can't pay in tenths, or hundredths of a cent. It's the same idea as a cryptocurrency ledger, minus the ridiculous overhead of the blockchain. Building out the infrastructure to do that is beneficial to both humans and autonomous agents.

https://x402.org/

So... yay?

This is all about leverage.

You don't wait until your one and only supplier of land decides you need to tithe 50% of your earnings for the privilege of setting up your business before you diversify.

Look at the Strait of Hormuz. The countries with the backup plan (rail built years ago) have an option other than dealing with Iran. The ones that don't suddenly have a problem - a big one.

Datacenters can indeed be built more cheaply on earth... for now. In the event that changes... you want to already have your alternatives tested and ready to be scaled up. Maybe this will all turn out to be just a waste of money that historians will later chalk up to ego.

To put it another way... if you wanted to test an untried technology and attempt to scale it up, would you rather do it while people are throwing buckets of money at you, or would you rather do it when your cash flow is under threat and investors are getting cold feet?

There's also this lawsuit:

https://www.syracuse.com/micro...

"A national advocacy group and some Central New York residents filed an 11th-hour lawsuit today seeking to block Micron Technologyâ(TM)s development of chip fabs in the town of Clay, arguing that the environmental review of the massive project was inadequate.

The lawsuit was filed the same day that state and federal officials joined Micron leaders for a long-awaited ground-breaking at the site.

The litigation was filed in state Supreme Court in Albany by Jobs to Move America, a national nonprofit, and Neighbors for a Better Micron, an informal group represented by Clay resident Bonita Siegel.

The lawsuit claims that the Onondaga County Industrial Development Agency failed to adequately consider the environmental impacts of the massive project before approving it in November."

Apparently "Jobs to Move America" is based out of California... so this seems to most closely match the OP's claim that "1 guy and 9 out-of-State activists shut down a planned Micron fab in New York."

However... the fab is still planned for construction, they've just de prioritized it in favor of fabs in Idaho, where presumably there are fewer encumbrances to getting shovels in the ground.

https://www.techpowerup.com/34...

"Micron has announced a significant revision to the schedule for its semiconductor campus near Clay, New York, with initial production now set to begin at the end of 2030. According to company filings and permit documents, the construction timeline for the first manufacturing facility has been extended from three years to four, and regulators have approved this adjusted schedule. Additionally, Micron has amended its $6.1 billion CHIPS Act agreement to reallocate approximately $1.2 billion from the Clay project to expedite expansion in Boise, Idaho, allowing the Idaho site to become operational well before the New York facility. This is part of Micron's strategy to maintain at least 40% of its DRAM manufacturing operations in the U.S.

The revised schedule significantly delays the following phases of the Clay campus and changes the project's long-term outlook. According to the new plan, the second fab is now set to begin in mid-2030 and is expected to be completed around 2034. Similar delays have affected the third and fourth fabs. As a result, the completion of the entire four-fab campus has been pushed from the original target of 2040 to closer to 2045. This delay also affects the introduction of community support initiatives, such as childcare, housing, and transit improvements, which are intended to benefit the campus workforce. Local officials attribute the slowdown primarily to labor shortages and extended construction cycles."

BTW, for those who want to know, the lawsuit was settled:

https://cnycentral.com/news/lo...

"CLAY, N.Y. â" The Onondaga County Industrial Development Agency on Thursday approved a settlement to pay Azalia King of Clay nearly $3 million to move out of her Caughdenoy Road home.

King's home is on land that tech giant Micron will need for its chip fabrication plant in the Town of Clay and so is a 6-acre parcel of land she owns across the street.

"There was no circumstance that we could have Mrs. King on that property and have Micron built around her," explained County Executive Ryan McMahon. He said the deal with King needed to get done, or the project could've been at risk."

Is this what you're referring to?

https://www.wired.com/story/mi...

"Micron Megafab Project Faces a New Hurdle as Activists Seek a Benefits Deal
Activists are demanding a way to hold the memory-chip maker accountable to its promises to protect the environment and embrace communities of color in central New York."

Or this?

https://www.wired.com/story/mi...

"Legal Battle

Across the US, governments reserve the power of eminent domain to seize real estate and redeploy it for a greater purpose in exchange for fair compensation to owners. Comprehensive data on how often it is used is lacking, but King has been a target more than perhaps most Americans.

Around 1965, Onondaga County used the threat of seizure to force King and her husband, Glenn, out of their farm to make way for a power station, according to court papers. Thatâ(TM)s how King ended up at her current residence on Caughdenoy Road, along the western border of Micronâ(TM)s project.

During the dotcom boom, the couple faced seven years of additional county pressure to sell their land to allow for a semiconductor fab, court papers show. In 2005, they relented. The Kings sold their 47-acre property to the county in exchange for $330,750 and a license to live tax-free on 3.61 acres of the land until both of them were dead. The envisioned fab never materialized. Kingâ(TM)s husband died in 2015.

Micron announced its New York project in October 2022, a megafab that would surpass the much-touted TSMC chip complex in Arizona. Current plans call for the first chips to ship in late 2030, about two years behind schedule. Displacing King has been a prerequisite. Her land is set to house parking garages and rain basins, project documents show.

In recent months, Onondaga County turned to state eviction and eminent domain laws to try to kick King out of her house no later than mid-January, under the threat of fines. Last Monday, King sued the county development agency in state court, contending that any forced move would upend her life and violate her lifetime contract.

King âoemerely wishes to live out her remaining years in her home, a place where she feels safe, comfortable and can have her family visit,â the lawsuit states, noting she has three dozen grandchildren or great-grandchildren. âoeDefendant is attempting to back out of the agreement ⦠simply because plaintiff has lived longer than defendant anticipated ⦠and the agreement has become inconvenient.â"

Neither of these stories indicates that the fab is dead... just delayed.

Who says they're not expanding?

https://www.benzinga.com/tradi...

"Micron expects fiscal fourth-quarter capital expenditures of around $10 billion, bringing total fiscal 2026 capital spending to approximately $27 billion. The investment pace isnâ(TM)t expected to slow anytime soon.

Chief Financial Officer Mark Murphy said the company expects quarterly capital expenditures in fiscal 2027 to exceed fiscal fourth-quarter levels as Micron accelerates construction of new clean-room capacity to meet long-term AI demand.

âoeWe expect quarterly CAPEX in fiscal 2027 to be above fiscal Q4 levels,â Murphy said, adding that more than half of the increase next year will come from construction spending as the company expands manufacturing capacity.

The investments include leading-edge DRAM fabs in Idaho and New York, continued expansion in Taiwan and Singapore, and additional advanced packaging capacity aimed at supporting next-generation high-bandwidth memory (HBM) products that power AI servers."

https://www.trendforce.com/new...

"[News] SK hynix Advances DRAM and NAND Roadmap, Targets 3x Wafer Output by 2034, 375-Layer NAND at Year-End

Nikkei reports that SK hynix is building four semiconductor fabs in Yongin, with the first facility expected to begin operations in early 2027. The original schedule had stretched to 2045, but the timeline has been pulled forward by roughly a decade, according to the report. Chairman Chey Tae-won added that demand is rising so rapidly that even the accelerated expansion may still fall short.

As previously reported by The Elec, SK hynix is targeting a significant ramp in DRAM output, lifting monthly wafer capacity from around 550,000 wafers today to roughly 1 million wafers by 2030. Much of the expansion is centered on the Yongin Semiconductor Cluster, where the first fab alone is expected to add about 360,000 wafers per month of DRAM capacity in the first half of 2030, according to the report."

https://www.datacenterdynamics...

"Memory chipmakers Samsung and SK Hynix are reportedly scaling up production of high-bandwidth memory (HBM) in order to meet growing demand from AI customers.

According to Korean news outlets, Samsung is looking to expand its production capacity by around 50 percent in 2026, while SK Hynix has announced plans to increase its investment in infrastructure by more than four times the figure previously announced.
Server memory
â" Thinkstock / NorGal

Both companies are currently constructing new fabs in South Korea to support these ambitions. Samsungâ(TM)s P5 facility in Pyeongtaek, Gyeonggi, is expected to be operational by 2028, with SK Hynixâ(TM)s M15X facility slated for utilization by mid-2027."

There's more than enough demand that not expanding production is just leaving tons of money on the table. You may legit be pissed that they're prioritizing HBM over DDR5 or DDR6, but money talks, and that's where the demand is.

We're not facing an OPEC situation... at least not with the manufacturers. Maybe with the distributors and the speculators who bought up all the ram and are ransoming it back to us a piece at a time at nosebleed prices?

I mean... this is just like farmers selling some of their yet to be harvested crop on the futures markets.

Sure, they could hold out for more, but something unexpected like a drought or a storm that damages their crops (or a war that shuts down the Hormuz strait) could happen.

In other words... it's risk management. Just like Southwest did for years with fuel hedging. They paid more up front to lock in prices for longer, and it paid off when oil spiked.

The customers want to guarantee a minimum supply even if prices go higher and are willing to lock in contracts to guarantee it. Micron gets money even if something unexpected happens (for example, China bans import of HBM to prop up their own domestic ram manufacturers, or the current administration bans sales of HBM to Chinese companies.)

I wouldn't read anything into this other than Micron's management seems to be acting rationally.

"which generates electricity without carbon emissions"

Ehh... Until they do resource extraction for uranium without fossil fuels, that's not quite true.

Still better than coal and natural gas fired plants, although all three generation types still require steam turbines.

Speaking of open vs. closed loop systems... how much water does a power plant (nuclear, coal, or otherwise) consume vs. a data center, during normal operation?

I listened to a radio interview with Bernie... apparently he wants this so that "citizens can block harmful policies". So that probably explains the 50% - they want voting shares in order to control the boards.

This is a bit more than just taking a stake in companies in order to give taxpayers an opportunity to share in the gains... this is basically nationalization without taking over the entire company.

So I guess I was wrong about where this was coming from. Too bad, since if the tech bros are right and this leads to the ultimate infinite money glitch, having a resource wealth fund that pays dividends would have been one way of softening the massive dislocation in labor allocation for both white and blue collar work.

Correct. Presumably if your investors don't like the idea of getting diluted by uncle sam, they'll cut off sales once you hit 199M a year.

You'd need to get tangible benefits for letting the government "buy in"... enough that your existing investors and share holding employees would be ok with crossing the 200M cutoff.

Probably a foreign held corporation in Ireland that owns all the IP and gets paid licensing revenue from all the sub 200M companies will serve as the controlling entity.

What's odd is you could accomplish something similar to Bernie's plan by funding AI startups and taking an equity stake. Upon IPO, you sell the stake and then take the money to go incubate more companies. The leftists complain about the massive imbalance of wealth, but are unwilling to go and take the risk that the VC funds are taking by backing a bunch of unproven companies that may never pay off.

Dilution is preferable to paying cash taxes on unrealized gains - the other socialist/progressive "solution" to people being successful. And, one would hope that negotiations would get things closer to a reasonable percentage from 50%, like the 10% the US government took in Intel.

It aligns the interests of the government with the company - you can't pay dividends on a wealth fund based on equity stakes if you take actions (policy or otherwise) that tank the value of the stock.

If anything, having the government as a partner (preferably a silent one) can help boost the stock price. If those shares aren't trading, then they aren't impacting daily price discovery. The only impact is that if the company ever pays dividends, half of the dividends go to uncle sam.

I'm not a proponent of nationalizing companies, but for AI companies to hit max velocity for spend, they need every advantage they can get to clear red tape and public opposition out of the way.

From an October 15th Pew Research report:

https://www.pewresearch.org/gl...

"But many are worried about AIâ(TM)s effects on daily life. A median of 34% of adults say they are more concerned than excited about the increased use of AI, while 42% are equally concerned and excited. A median of 16% are more excited than concerned.

Concerns about AI are especially common in the United States, Italy, Australia, Brazil and Greece, where about half of adults say they are more concerned than excited. But as few as 16% in South Korea are mainly concerned about the prospect of AI in their lives."

Contrast this to Biden era view on AI - which was to legislate regulations and restrictions on development which most likely would have been only to the benefit of the largest players. Open source players in this space would remain unaffected, since the metric is 200M in sales.

That Bernie and Trump are talking similar language (US government taking a stake in AI companies) makes this at least a starting point in discussions. People talking about UBI without proposing a mechanism for financing UBI - this is one possible way of financing things.