Pete Schmidt

What stands out to me is that Charleston, SC, is not a center of economic activity, despite being one of the best ports in the colonies through today. Atlanta, GA, is, though, and it's got no port at all.... (4/end)

I did an overlay of the two images. Only two areas of high economic activity didn't have railroads: Phoenix (railroads reached Phoenix in 1895) and Miami, which didn't have a rail link until the Flagler railroad in 1912. (3)

The economic activity aligns with this 1889 map of US railroads. (2)

This is a great map. However, we can make this more interesting. (1)

Data source: mpdc.dc.gov/publication/...

In 2014, the DC Metropolitan police made 20,387 arrests, or an average of 55 per day. At that rate, they would have made 726 arrests in 13 days, or a 32% more than 550

A cubic meter of the sun’s core produces 6,360 megawatts, approximately the same as the Grand Coulee Dam’s maximum output of 6,800 megawatts.

(The sun’s core produces 3.86E26 watts within a radius of 139,000 km, per Wikipedia; Grand Coulee dam stats via USBR.gov.)

The problem with this is that toxicity is mostly about off-target effects (with some target overload and receptor affinity concerns) and to evaluate off-target effects, you need a whole organism. Simulation and organoids are useful when you want to consider a specific target.

Within human intelligence, processing ambiguous situations add approximately 600ms to processing time, and adds activity in the DLPFC and OFC. Maybe without those 600ms, our brains are actually processing an input-output transformation, like an LLM. Maybe those 600ms are where intelligence lies. (6)

As we marvel at the progression of the technologies we group together under the heading AI, we are increasingly unable to distinguish between a sufficiently sophisticated transformation of input to output data, and the kind of general-purpose processing that characterizes human intelligence. (5)

If LLMs are information theory constructs, then doubling the rank of the tensor weights would quadruple the context window. Doubling a log is equivalent to squaring the message length. Continued progress could easily extend beyond our ability to recognize entropy limits in interactions. (4)

She Is in Love With ChatGPT A 28-year-old woman with a busy social life spends hours on end talking to her A.I. boyfriend for advice and consolation. And yes, they do have sex.

ChatGPT has a "context window" which was described as 30,000 words in this 15/Jan/25 article. This makes sense in terms of an information theory entropy limit. It doesn't make a lot of sense when we think about traditional definitions of intelligence. (3) www.nytimes.com/2025/01/15/t...

With a robust enough transformation tensor, information theory suggests that an input could be transformed to an arbitrary output without the tensor describing what people would consider "intelligence." Information theory offers an explanation of how the Turing test can be wrong. (2)

Image showing an excerpt from Claude Shannon's treatise on information theory.

We think of LLMs in terms of "artificial intelligence" but there is another, more accurate way to think about them: in terms of information theory. Information theory tells us that the information content of messages (Shannon's "entropy") is proportional to the log of the message length. (1)

In information theory, the information content of a message increases as the log of its length. This suggests that a doubled size array of LLM tensor weights would allow you to square the prompt length. If an LLM can respond to a 100 word prompt, an LLM 2x the size could handle a 10,000 word prompt.

In 2007/8, I worked with genetics labs while Bush had the “no fetal tissue” stem cell ban. In university labs, they had 2 of each piece of equipment: 1 for general studies and 1 for federal grants - no fetal tissue. They charged that 2nd set on direct budgets, because it wasn’t shared. (5/end)

This means the choice is: direct charge the equipment: $600k direct + 15% indirect. Indirect charge the equipment: $100k direct + 60% indirect. So that’s $690k vs $160k. And from a pure numbers perspective and 280 characters of insight, paying $690k sounds like the better deal. 15% overhead! (4)

Second, it actually under-pays for infrastructure. Imagine that scientists need a $500k psychofrackulator to do their $100k study. If it is used only for the study, they can charge it as a direct expense, for $600k. If they share it with even 1 other study, they can only cover it as an indirect. (3)

This is really very advantageous to the government, for two reasons. First, many federal grants require an in-kind contribution of capabilities and services. I worked on a $10M/year grant to UMiami in the aughts and the terms of the award required a lot of in-kind services. (2)

There is some confusion about how percentages and fractions work that has contributed to confusion about indirect rates in federal awards. For example, a 60% indirect rate means that the indirect/overhead charges are calculated to be 60% of the direct charges. This is 38% of the total. (1)

What is your position on Van Gogh? I was embarrassed when I first heard a Dutch person say the name, but then I decided to relax and not worry about it. Verbalization is for communication, so if “the right way” is unintelligible to your audience, it is wrong.

For all the people who think that they need to be exposed to virulent pathogens to “exercise” their immune systems: Do you think that your immune system is not constantly destroying bacteria and viruses even when you have no symptoms?

Santos is out. It’s the end of a fabulous era.

Oops, I meant “fabulist.”

There seems to be a Dunning-Kruger aspect to SBF: smart enough that people convinced him that he was smarter. But to not understand the game theory that underlay his strategy means he wasn't all that good at math. The game theory he describes in interviews shows his thin grasp of strategy. (5/end)

SBF/FTX ran this way - understanding odds x return = EV, but not that a bet with a 0% chance to win infinity dollars can have a positive EV but be stupid to play. FTX made a bunch of massive upside vs implode bets. They won massive upside several times. But they imploded once. That's enough. (4)

Play 25 games with a 51% chance of +1 & 49% chance of -1. Net expected value (EV) is 1.5, or 25 times 0.02 net positive. Now run 25 games with a 51% chance of 2x return and 49% chance of ending the game. EV = 1.64, because 2^25 is a big number. But the odds of a non-zero return are 1 in 2x10^7. (3)

Why is this bad math? After two such wagers, the odds of destroying humanity would be 74%. SBF’s analysis ignored that his model had negative but no positive game-ending outcomes. The right framework for such a thought experiment is not “double or nothing,” it’s “double or half.” (2)

SBF is supposed to be a math genius; I don’t buy it. Why? He repeatedly expressed enthusiasm for “double or nothing” bets: he has said he’d take a bet on a 51% chance to double the viability of humanity even against a 49% chance of destroying it. This is bad math. (1) blog.shrm.org/blog/sam-ban...

The two worst pandemics of the modern era, the 1918 H1N1 and the 2020 SARS-CoV-2 pandemics, were caused by viruses in their first jump from animals to humans. Many other epidemics share this characteristic.

Call your congressional offices (1 rep + 2 senators) and tell them that you want NIAID to continue to advance our knowledge of infectious disease (after a pandemic!) and immunology. Many recent breakthroughs across medicine have come from new insight into antibodies and immune system cells.