Don't be a hassler — "people who regularly cause problems or make life difficult". Stressful People in Your Life Could Be Adding Months To Your Biological Age science.slashdot.org/story/26/02/...

Make sure you are registered to vote well before the election.

For the AI doomers in the house, worried that a new technology is crashing employment…we have been here before. It was fraud last time. What is it this time? Long Before Tech CEOs Turned To Layoffs To Cover AI Expenses, There Was WorldCom tech.slashdot.org/story/26/02/...

War is hell So is coal Solar shines on Ukraine bsky.app/profile/prof...

So annoying when you discover an airline quoting different prices depending on which browser/device you use to book.

“One mile on a bike is a $.42 economic gain to society, one mile driving is a $.20 loss.” “Which means that Copenhagen, a city of 1.2 million people, saves $357 million a year on health costs because something like 80% of its population commutes by bike.” #CityMakingMath Some costs aren’t costs.

Good morning with good news: Solar, wind & batteries will be 93% of US record-high new utility-scale capacity additions in 2026! Gas additions of 6.3 GW will be less than 6.4 GW of gas & coal retirements! Solar adds will be ~50 GW, with small solar included. www.eia.gov/todayinenerg... #energysky

Absolute shocker of a rights grab - academics may want to push back hard

The Economist is seeking promising journalists and would-be journalists to apply for the 2026 Richard Casement internship in London. Apply by March 6th and find out more here:

Iris 7 #americanart #georgiaokeeffe

NEW | India reached 50% clean power capacity 5 yrs ahead of target, reflecting its accelerating #energytransition – but state-level progress is uneven 🇮🇳 Our report with IEEFA evaluates the progress across 21 Indian states👇 ember-energy.org/lat...#energytransition #SET2026

Paulie ball!

WindWaterSolar met a peak of 144% of demand on the California ISO grid on Saturday 2/21 Demand down 1.1% so far in '26 v '25 and down 1.7% v '23 (despite growth in EVs, heat pumps, data centers) Fossil gas down 35% v '25 and 54% v '23. Batteries up 39% v '25 and 304% v '23

And why bother with new nuclear when enhanced geothermal does the exact same thing but faster, cheaper, cleaner & safer and with less land & no mining and can be done most anywhere web.stanford.edu/group/efmh/j... news.stanford.edu/stories/2026... Nuc pushers have no idea what the alternatives are

So much that’s good in this, but clipped out this in particular. Yes, yes, yes: it’s a losers’ ideology, demonstrably, historically so. It’s stupid and pointless and wearying, and, as he says, it ends up eating itself. It’s who it takes with it on the way down the shitter that’s the tragedy.

Capital costs of new gas plants now are more than capital costs of solar plus battery. Gas plants also have large, volatile fuel costs to operate. Solar & battery plants have zero fuel cost. Wind & battery plant often has total costs less than gas. emp.lbl.gov/sites/defaul... #energysky

Price of large gas plants more than tripled from ~$1,000/kW to $3,587/kW at proposed Ohio plant! Then there's the additional, volatile gas cost to operate it! Solar plus battery or wind plus battery capital cost are much less & with fuel cost of zero!! www.bloomberg.com/news/article... #energysky

To be clear, I am no expert either. But for 30 yrs I have been thinking about how models of neural systems represent statistical information, since I worked on how fly brains process visual signals. Stat mech places strong limits on how LLMs can work despite grandiose tech bro claims otherwise.

And evidently this function is *very* poorly coded, b/c the quantitative result is that the entropy of the ensemble of output passwords is several fold smaller than the entropy of the training set. So much better to just use a random number generator, b/c while imperfect it is much less imperfect.

The training data is by definition finite and you would in fact hope that it does not include all possible passwords. Then, because the LLM is just statistically sampling and recapitulating the training data, it outputs a very small sample of possible passwords.

I find it is useful to think about training the LLM as a signal processing problem. You have a finite set of signals in the form of all the time (or other) series in the training data. These are transmitted via a noisy process to the output of the LLM. S(output) must be ≤ S(training set).

But if that is the relevant definition wouldn't the entropy of the training set for a modern LLM be large enough to *in principle* produce good passwords? So either way, I don't see how "entropy of the training set" per se excludes use as a password generator. But I'm definitely not an expert!

Because I always forget the keystrokes for inequality symbols, I find this site useful for rapid cut and paste. Inequality Symbols symbolsdb.com/inequality-s...

As corroboration of this @canarymedia.com report, here is a preview of a thing I am getting ready to publish. Despite TX wind + PV increasing from ~9% to ~33% of supply from 2014 to 2024, customer outage min were basically unchanged, and MED assoc outages were caused by fossil fuel failures. 🔌💡

Of particular relevance in the context of all the recent authoritative reports claiming LLMs can help novices develop and build threats. #biosky

Very good representation of how that "crinoid mouth" that's really an angled (rounded?) slice through a set of crinoid stalk segments works. The "slice" is an erosion related thing that's probably pretty recent (caused by the formation of the little rock clast the crinoid is preserved in).

i procrastinated work to ask @echinoblog.bsky.social questions about that "smiling rock" fossil and ended up with this after everyone's verbal explanation. lots of people seemed confused so maybe this helps. (see the entire thread for context)

Of particular relevance in the context of all the recent authoritative reports claiming LLMs can help novices develop and build threats. #biosky

🚨The most important graph in the world has another year of data, and it doesn't look good!🚨 NASA CERES satellites measure how much sunlight Earth absorbs and how much heat is radiated back to space. The difference is heat accumulation, which has more than tripled!

As corroboration of this @canarymedia.com report, here is a preview of a thing I am getting ready to publish. Despite TX wind + PV increasing from ~9% to ~33% of supply from 2014 to 2024, customer outage min were basically unchanged, and MED assoc outages were caused by fossil fuel failures. 🔌💡

This is another interesting study, Oregon centric, recently updated (draft): washingtoncounty.civicweb.net/document/321...

Preliminary studies, priority for the region is more frequent service and replacing extremely old rolling stock, HSR planning. But, they did get an electric bus for the amtrak service :)

The entropy of the ensemble of LLM outputs, from a given LLM trained on a particular data set, cannot be larger than the entropy of the training set.

Just took a train PDX to SEA, have been standing next to King St Station waiting for a bus. Very Loud. Who has done a study of the cost/energy/health savings and overall quality of life improvements of electrifying US freight and passenger rail? #energysky

The information/entropy content of the training set is your favorite version of sum[(p)*log(p)] over all the transition probabilities of tokens of all sizes in the set, or a Kolmorgorov calculation of the minimum number of instructions necessary to replicate same.

Welfare Queen in TWO Cadillacs. For those who are old enough to remember. Coast Guard Reportedly Spent $172 Million On Private Jets For Kristi Noem’s Use www.forbes.com/sites/sarado...

Imagine being a Bobby and being informed this is your duty assignment for the day.

HOW MANY TIMES DO I HAVE TO YELL THAT THE ENTROPY OF LLM OUTPUTS CANNOT BE LARGER THAN THE ENTROPY OF THE TRAINING SET!!!????!!! This is even an application in which entropy can be added, but evidently none of the coders knows enough information theory to bother. www.theregister.com/2026/02/18/g...

A MEETING IN THE TWILIGHT SKY: Last night, the whisker-thin crescent Moon had a rare close encounter with Mercury. Dr. Gilbert Plumer photographed the conjunction from Rockledge, Florida spaceweathergallery2.com/indiv_upload... spaceweather.com

Man, the Amtrak trainsets running between Portland and Seattle are *super* rickety.

2026. Italy’s Iconic Lovers’ Arch Collapsed—Of All Days—On Valentine’s Day gizmodo.com/italys-iconi...

27th of 49 days in 2026 with >100% WindWaterSolar on the CAISO grid. Gas own 54% in vs 2024 and 56% vs 2023 Batteries up 54% vs 2024 and 301% vs 2023 PV up 38% vs 2024 and 56% vs 2023 PV+Batteries displacing gas Imports up too: These are ~58% WWS

This is the way. Each Russian demand list must be mirrored. The battlefield is stalled and no one is winning. Thus, there's no reason to preemptively concede a single russian demand. The deference to russia's supposed might, which has long characterized European/US policy needs to end. Russia...

👇🏼

Good morning with good news: Battery costs plunged 27% in 2025, a bigger fall than 11% forecasted 2025 decline! Global battery deployment jumps 33% in 2026 to 122.5 GW in 2026! Battery costs fall 25%; solar costs 30%; onshore wind 23% & offshore wind 20% by 2035! www.bloomberg.com/news/article...

Yay! New orca in town.

Yup. Apple’s Lack of Interest in AI Hype Is Suddenly Making It Attractive to Wall Street gizmodo.com/apples-lack-...

“But thank you.” Class.