ckrapu

Anecdotally, a prestigious consulting firm (one of McKinsey/Bain/BCG) essentially stopped hiring MBAs and instead hired several friends from my PhD cohort despite the B-School ranking hire than most of our graduate programs.

I've always wondered what signal they were acting on. Perhaps the value of the MBA has been watered down, or it was just too easy to game the admissions.

> Being forced to survive an alien landscape can make you whole again.

I can't agree more. My sister died at the hardest point of my PhD and I buried myself in my work for nearly every waking second for years, confident that at least I was doing it for myself. I couldn't have done that if I were working a normal job.

The PhD is has become the de facto replacement for the advanced workforce training programs (apprenticeship, guilds, corporate talent development programs) that many civilizations used to use. For some fields, you really do need to bang your head against a small number of problems without anyone holding your hand before you become proficient.

Not all PhD graduates get there; many just skate by because no one wants to fail them. They are an essential part of the modern labor force, though.

In the autoregressive decoding framework, the hidden layers' state for computation of token `t` is conditionally independent of all hidden states for `t-1`, `t-2` and so on given the observed tokens.

Put differently, the observed tokens are a bottleneck on the information that can be communicated across tokens. Any scheming performed by an LLM which requires more than one token to formulate must therefore pass through the visible tokens. With opaque vectors transferred across decoding steps, this is not the case.

The computation in the hidden layers, as far as we can tell, is not sufficient for scheming in a single decoding step. It looks like it requires O(10^2) or O(10^3) steps instead, judging from anecdotal evidence like the reports of scheming from o1 (https://cdn.openai.com/o1-system-card-20241205.pdf)

As far as your last point goes, I'd rather have a more transparent system, all other factors held constant.

Identifying scheming in the latent streams would be harder as you would have an extra layer of obfuscation between you and the model’s reasoning.

My opinion is that opaque reasoning is a prerequisite for many of the worst possible AI outcomes.

We should make reasoning fully visible in the output space.

> Then explain why if it works so well why Musk paid $10B in taxes when he sold Tesla stock if it works so well.

Loans to Musk were a risky asset before the election. He was at risk of investigation in the event Trump lost. No one would have given him a loan big enough to supply the cash needs which this sale supplied.

That said, ordinary billionaires can of course make use of this strategy.

There is a wide range of moral and practical opinions between the statement “all weapons are evil” and “global corporations ought not to develop autonomous weapons”.

From an AI risk perspective, one of the most wonderful things about LLMs is that their chain of thought can be entirely read off their outputs by humans with no specific training.

This is a risky step backwards, and for apparently little gain.

I left geospatial for big tech. It wasn’t because of the pay. It was because my colleagues with GIS and geography backgrounds were infuriatingly mismatched to the data science / ML projects they were tasked with. They also had trouble admitting when this was the case.

That does seem like a much better approach than freezing it, then.

Exactly.

I wasn't satisfied with his idea because it assumed that the biomass had to stay dry indefinitely, despite being locked behind an impermeable barrier. It's also not very convenient geographically, because raising crops (at least for low-value types like switchgrass) is not economically viable in the driest areas where they would be stored.

I think it's more like 3-4 orders of magnitude (i.e. 1000 or 10000 of these sites), but yes, it would take quite a few to completely offset CO2 emissions.

Any idea is enlivened by the possibility of criticality and runaway thermal energy production as the melting hay produces heat, thus melting more hay and so on.

I do this with Cursor for research work

I was hoping that this proposal would be a uniquely unilateral one. If a single state or province in the USA or Canada committed to this plan (and if it worked as well as I optimistically propose, and provided enough nutrients) then they could singlehandedly put away enough carbon to actually solve the problem.

The spring-summer timeframe is the biggest issue to deal with here. I think that this design would have to have a sacrificial layer with decomposing, anoxic material which degrades during the warmer months. Perhaps it could be mitigated with various expedients to decrease the ambient temperature like flooding the surrounding landscape or putting a shade over the pile.

This scheme only makes sense if the amount of biomass added every year is much larger than the amount which is present in this external layer.

Anecdotally, my father has told me about soaked hay bales out in the pasture which still had a frozen interior core by late May.

Good catch :)

I thought about this for awhile and the gap between harvest time for many crops and first frost isn’t enough to get more than a few inches of rainfall in most agricultural regions with favorable economics.

I think the wetness will wreck the insulation of the first meter or so, but won’t lead to much convective heat transfer if the outside never gets saturated. A big if, to be sure.

As an aside, it’s common practice to leave large piles of grain outside overwinter in the central USA and it’s not optimal, but they certainly don’t saturate the whole way through with water.

Let's say that we have a hollowed-out zone in the middle of the biomass pile where we tolerate limited oxidization so we can run a fire. If the rest of it is wet, maybe the heat from that combustion could pyrolyze a large radius of surrounding material since O2 flow into the system should be small.

A flawed doomsday weapon but a good mechanism for building a fictional world where the biosphere develops underground.

One downside: you have to ship all the carbon to the coast. Transport is a non-negligible consideration for all of this. Ideally, you just grow a ton of switchgrass in northern US / Canada / Siberia and store it nearby.

I had an issue with an out-of-date mathjax configuration, sorry! I've updated it.

Is it still not rendering? It looks fine in my browser.

Great point. I handwave this away by saying that most of the biomass will be frozen most of the time, providing the necessary structural support.

This is definitely true.

My hope would be that the thawed region would be a very thin shell overall, so the overall emissions as a fraction of total stored mass would be relatively low. Can you think of any ways to minimize anoxic activity in the thermally active area?

Sadly, I don't think so. Many of these carbon burial/sequestration proposals all advocate just taking all of the plant matter and tucking it away, including the N and P.

What’s the tasking time on these? Could future governments use these for military purposes?

From anecdotes I've heard and read, the labor practices used in Taiwan in the semiconductor industry clearly were not going to be legal in the USA. Did TSMC think that they were just going to be able to flout regulations and go around the unions?

I grew up next to Fargo and always wondered what it was like to use their product.

Do you have any stories or anecdotes to share?

The guy’s twitter bio says he proved P!=NP. Clearly a crank.