strangeloops85

A very important point! In Tristan (NYU prof)’s write up he noted evidence of the OpenAI mathematician team doing a lot of correction and guidance along the way. We are never told about this with openness and clarity. At a minimum, complete disclosure and honesty is needed by the companies and about the precise role of their staff members.

Are these companies expending any of their lobbying efforts or funds on this problem? Or are they just going to the easiest option? I'm not shedding any tears for the AI companies and their infra challenges. They're the ones driving a huge delta increase in energy use and emissions this coming decade, so their actions merit scrutiny and they are responsible for their actions and the resulting emissions that we will all suffer from. If I had my way, we'd stack an enormous carbon tax on all of them to force their hand.

Look, these companies are the ones selling a story of being benevolent (notwithstanding AI killing us all or putting everyone out of a job). And Anthropic is explicitly a B-corp. How does any of that square with taking any compute they can at any environmental cost?

The Memphis case to me is very illuminating: https://www.france24.com/en/live-news/20260811-memphis-ai-da...

But it's ok, they're EAs and gave $1M to a Memphis nonprofit on 'AI education', while spending a $1B/ month renting Colossus:

https://www.fox13memphis.com/news/anthropic-gifts-memphis-no...

Gotta love how all the companies' platitudes about sustainability become ludicrous as they begin burning natural gas at prodigious scales. But don't worry, "AI will solve climate change" so we can just pump extra CO2 now, right guys?

I am deeply, deeply cynical and honestly angry about these companies and their self-righteous protestations of being ethical, or public benefit corporations or whatever. Anthropic is benefiting from xAI's Memphis data center that is violating (previous, not anymore!) EPA standards, causing disproportionate impacts to nearby African-American communities. All the while, Ants appear to be floating on some higher effective altruist / B-corp plane. The little people and their petty concerns are just bumps on the road to RSI and ASI after all.

The energy economics in space are also a bit more complicated than usually thought. I think Starlink has been using Si cells instead of III-V-based ones, but in addition to lower output they also tend to degrade faster under radiation. I guess that's ok if the GPU is going to be toast in a few years anyway so you might as well de-orbit the whole thing. But that same solar cell on Earth will happily be producing for 40+ years.

Also the same issue with radiative cooling pops up for space solar cells - they tend to run way hotter than on Earth and that lowers their efficiency relative to what you could get terrestrially.

The elephant in the room for all lunar scenarios is lunar regolith. Even ignoring the toxicity to humans (big problem and will happen quite quickly for any humans there!), it will be a big long-term problem for robots and machinery in general.

Really outrageous extortion. UCLA and UC as a whole are running budget deficits and facing reduced state support (as well as the grant cancelations). There's simply no money for this and the white house knows it.

The particular type of fraud described here (paper mills etc.) is less common in the U.S. (different types of fraud may exist but that's more subtle and complex). There tend to be specific geographic clusters associated with this behavior that have to do with how university expansions have been done in many countries.

Oddly enough, pre-LLMs, I would have said most of these crap paper mill papers didn't really affect the actual fields. Yes, they cited each other but outside the citation ring didn't really alter the field in a knowledge sense. But now.. if these get picked up in Deep Research it's a problem.

It’s interesting how hard and widespread a push they’re making in advertising this - at this particular moment, when there are rumors of more high level recruitment attempts / successes by Zuck. OpenAI is certainly a master at trying to drive narratives. (Independent of the actual significance / advance here). Sorry, there are too many hundreds of billions of dollars involved to not be a bit cautious and wary of claims being pushed this hard.

I assume there was tool use in the fine tuning?

To those saying all you need is a lot of radiators.. remember that the radiators themselves gain heat from both sunlight and the Earth itself. It is a surprisingly tricky problem and, yes, all heat can be dissipated to achieve a desirable set point given sufficiently large area. But it is certainly not easier than just having say an economizer and dropping the data center in Iceland or a cold place. Makes no sense

At least for many engineering and science disciplines a solution remains to force students to actually try to learn the material: in-class, closed book exams. Nothing is ideal, but this does force students to actually engage with the material and problems. They are welcome to use LLMs all they want to help them study (though they should be careful given how often I catch them making horrible mistakes in my discipline). But the assessment will be of them and their brain.

Public schools in the US get a relatively small fraction of their budget from state funding. The distinction between public and private is not as large or substantial as one might imagine.

For example the 10-campus UC system's total budget is $54 billion of which $4.6 billion comes directly from the state's general fund. https://lao.ca.gov/Publications/Report/4998 - the federal funding here is the same as for private universities, to do research or other work in the form of contracts/ grants.

I have to wonder what DOGE (including the alleged wunderkind Farritor that all the SV VCs were hailing) is planning for NSF. It’s amazing the arrogance of these people, to walk in and just do “hulk smash” on decades of hard work, infrastructure and institutional capacity.

This is perhaps a simplistic observation, but Bhartrihari’s linguistic idealism connects with how LLMs and their apparent intelligence is fundamentally language-driven. He would argue that is sufficient since our reality is fundamentally linguistic in nature:

“ Bhartrihari tells us, ordinary language is a crystallization of something already implicit in reality. Reality itself is fundamentally linguistic, and what we think of as language—ordinary language, with its words and conceptual divisions—is just a devolution or fragmentation of this more primordial linguistic totality. This is precisely why, for Bhartrihari, the ultimate reality is shabdabrahman, a linguistic absolute. So this is a strong form of idealism: things in our experience, and all things in existence, are fundamentally linguistic. We have no access to anything outside of language and therefore no reason to assume that there is, or ever was, anything separate from it.”

In theory it is less at risk, but in practice there may be fewer TAships due to general budget shortfalls and also more students competing for those spots.

OP here: I think the reason for reducing Ph.D. admissions is very simple and should be understandable to anyone who has ever been responsible for making payroll. We (at universities) have great uncertainty about future "revenue" (grants) with even funding for ongoing contracts/ grants not being guaranteed to come in next fiscal year. So we need to reduce expenses which are placed on the grants, the largest amount of which is paying for our trainees. The vast majority of universities in the US do not have extremely large endowments, and at least at the school I work at, the (very modest) endowment amounts that can be used for ongoing expenses already are.

I, as a PI, am not directly admitting anyone into my group this year to ensure I have enough funding to pay existing group members. We're hunkering down and making sure those we have now will be funded through the rest of their Ph.D. While this article is talking about program-level decisions, there is a bottom-up aspect as well - at my program and many others, we (faculty) directly admit students into our group and are often responsible for their salaries from day one. Many faculty are, at an individual level, making the same decision I am, to reduce or eliminate any admissions offers this year.

Edit: For reference, I am not at UPenn, but at a "typical" state school engineering program.

A very important statement is in the peer review file that everyone should read:

"The editorial team wishes to point out that the results in this manuscript do not represent evidence for the presence of Majorana zero modes in the reported devices. The work is published for introducing a device architecture that might enable fusion experiments using future Majorana zero modes."

https://static-content.springer.com/esm/art%3A10.1038%2Fs415...

Letting go of all probationary employees (hired in the last year) is precisely the wrong way to do any of this.

I suspect we will also be hearing about the long time experts that are the only ones who maintain or know about certain systems being rehired desperately..

DOGE is frankly ruining the reputations of Silicon Valley and software engineers in a way I didn’t think possible.. they’re being very stupid about what they’re doing.

Political appointees including the IRS commissioner have never had direct access to the system before given its sensitivity. There are likely legal ramifications that will be worked through the courts on this.

So the White House (where doge sits) has direct access to anyone’s tax returns. I assume this can be used to blackmail or threaten anyone they don’t like?

Imagine the outcry if the Biden White House had tried to do this, especially because Trump has never (to this day!) posted his tax returns..

If you're having a team going into every bit of government data, you might want to vet them. Beyond eugenicists and racists, if they're getting anywhere close to SCIFs, make sure they fill out that SF-86 accurately. And maybe even if they're not, have them go through regular clearance procedures.

Yes, it takes time. But right now, DOGE seems like a huge security risk for the USG what with it setting up backdoors everywhere, and with unvetted staff working for it.

Independent of whether these goals are appropriate or not, these actions (and those of so many other companies) are just so predictably craven.

These executive orders (and what "DEIA" exactly means or constitutes, legally speaking) have not been litigated or clarified yet. Is Google going to avoid interviewing anyone from a HBCU now?

At least Costco seems to have a logical reason for what they do and stood by it.

Have Elon Musk and all these engineers sworn oaths as required to be a federal employee? https://federalnewsnetwork.com/commentary/2019/10/the-oath-o... If Elon is "just advising", under what statute or directive does he have specific authority?

If they are accessing TS/SCI information and places like SCIFs have they filled out their SF-86? Are any of them dual nationals and do they have any ties or vulnerabilities to hostile foreign states?

Basic questions given the enormous access they are being given, far beyond frankly any handful of people have generally had in US government history.

Also, they have apparently plugged in their own private server at OPM. Has this already been compromised by Chinese/ Russian agents? Has the NSA had a look?

Who knows? The executive orders read like they were written by children and don't clearly define what they mean by "DEIA". But NSF's authorization is from Congress. Unless congress passes a law rescinding this as a part of what counts as broader impacts, or the Supreme Court rules that increasing participation of underrepresented groups is unconstitutional (by precedent it is certainly not!), then NSF cannot simply change the definition of broader impacts.

While everyone is focusing on the so-called 'DEI' part, please note that per this article they are also targeting "environmentally friendly technologies". This is incredibly broad and could de-fund a broad array of fundamental research that has, as potential long-term impacts, reducing carbon emissions, PFAS remediation, desalination, etc. etc. Given that proposals were written years ago, researchers did not know that in the new regime certain words would become verboten and may have framed the eventual impact of their work in entirely reasonable ways that are now considered ideologically unacceptable.

If they do choose to issue stop work orders on already funded grants that will certainly cause chaos across engineering and science in this country. And frankly, it will cede leadership on many topics to China (which is already dominating fields like materials science).

This is disingenuous: there is no "DEI" component of NSF proposals. There is a Broader Impacts section which is mandated by law - the America Competes Act of 2010, which congress passed. https://www.congress.gov/111/plaws/publ358/PLAW-111publ358.p... It states that:

GOALS.—The Foundation shall apply a Broader Impacts Review Criterion to achieve the following goals: (1) Increased economic competitiveness of the United States. (2) Development of a globally competitive STEM workforce. (3) Increased participation of women and underrepresented minorities in STEM. (4) Increased partnerships between academia and industry. (5) Improved pre-K–12 STEM education and teacher development. (6) Improved undergraduate STEM education. (7) Increased public scientific literacy. (8) Increased national security.

That's it. I've won many NSF proposals and have never talked about DEI. Instead we talk about outreach work we do with local schools, our involving undergraduate students in research who would not otherwise be able to volunteer their time, and of course the economic impacts of working on these topics.

An executive order cannot override the law authorizing the National Science Foundation and its activities. We are, for now, a country of laws.

The fact that you can pay to have your replies higher (plus the suppression of posts with links) on X are part of this. The other reality is that for me and my community / friends who used to be active on Twitter.. they're just not participating anymore.

It's become like Facebook became a decade or so ago. Total ghost-town with only some eccentric folks still posting. In the case of X, everything I see seems to be connected with some kind of ecosystem around Elon, Trump and right-wing influencers. Time to move on.

$10M/ year seems unusually high, even for med schools - are you sure those are annualized costs? Perhaps it's because I'm familiar with the more basic science (non clinical trial) parts of NIH, but a typical R01 is $250K-$500K/ year in direct costs so that would imply having to hold ~15 of them (with overheard) to reach $10M/year.

It seems like $10M in total value of active grants in a given year (~5 R01s) might be more typical (though the standard for tenure I've heard of in med schools is maybe ~2-3 R01s, so 5 would make you top tier and highly sought after).

For clarity, when I'm referring to $1M/year in annual funding, that's coming from ~6-7 active grants whose total value might be in the $5M range total.

$1M / year is on the upper end in the engineering school I'm at (average faculty brings in less than half that). I mostly have Ph.D. students (as do most people in my field) that cost ~$100K/ year incl. overhead, or postdocs that cost ~$150K/ year incl. overhead. Things are very different in medical schools where staff scientists are more common and also you have to fund your own salary.

Total household income is > $450K/year, which along with lots of saving early in our career, allows us to live in a home close to work and in a nice neighborhood. Everyone's experience will be different, and we are certainly lucky, but I am in a pretty similar situation to a dozen other recently-tenured faculty I know in my career/life stage at my university. We work hard, but the differences in hours etc. relative to my friends working at various Bay Area companies seem modest. YMMV.

Yes. Very expensive. h-index > 35. It varies. Both. ~$0.5-1M/year. Sure. Yes, and vice versa for my spouse as well.