I thought that The Goldfinch had a contrived beginning and a weak ending, but a long, delectable middle that made up for both ends. Also try the author's earlier novel The Secret History if you haven't read it yet.
It seems like a stronger story for robotics, since smaller models can always react to the environment faster than large models at a given hardware budget. Also because robots that keep their models local for latency or reliability aren't going to be carrying many kilowatts of inference capacity.
This reminds me of the decline of the news industry. The best thing that traditional news publishers offered was in-depth investigative reporting. For example, 50 years ago:
You can pick just about any year and find great work.
The best newspaper work is hard to sell. Pure investigative reporting is too expensive and slow to keep a newspaper in business by itself. Newspapers relied on geographically limited oligopoly to sell a lot of other cheap page-filler at good margins: weather forecasts, sports scores, political opinions, comics, classified ads, movie schedules...
People who just wanted to read the latest Calvin and Hobbes strip were inadvertently, benevolently financing the investigative reporting that their area's newspapers used to support, even if they never actually read those stories. When the Web un-bundled it all, everything but the investigative reporting could be done better and/or less expensively by non-news organizations. News junkies who really care about hard hitting news aren't numerous enough to keep funding the investigative reporting. (And even if they were, most metro areas don't have enough latent scandals to keep newspapers printing regular editions of investigative exposés alone.)
The weights are just numbers. I don't what technical background you have in other areas of computing, but I think that this is a good, short introduction that doesn't assume too much:
To quote part of it, Training a model can be thought of as tuning the dials on a really big machine. The way that a language model behaves is entirely determined by these many different continuous values, usually called parameters or weights.
Longer and slightly more technical, "Intro to Large Language Models" by Andrej Karpathy:
The trained weights are open, the training software is open, and the data that goes into training the model is open.
Not many models are fully open.
An open weights model is one that has freely available trained weights, and maybe fine-tuning tools, but it lacks the original training data (and usually lacks the training software). These are the most commonly used local models, like Google's Gemma series, Meta's Llama, or Alibaba's Qwen.
The first Idaho project is starting soon: "Micron has already achieved key construction milestones on its first Idaho fab with DRAM output scheduled to begin in 2027."
It's the striker strip that can be used in producing meth. The strip contains red phosphorus. Strike anywhere matches are the only kind that don't need a special strip. The strike anywhere kind were probably restricted due to their sensitivity to shock and friction which makes them more useful and more dangerous than safety matches.
There's not a big pool of well-adjusted people who are equally willing to learn the intricacies of Wikipedia and manage it better. The people who care a lot, in the wrong ways, outnumber people who are passionately neutral. Most people don't care enough to fight bad edits and reversions. They just stop contributing. I know I did. (I wasn't even editing controversial areas, just adding to data about chemical compounds.)
I still love Wikipedia, but mostly as a starting point to find deeper references. (Which, to be fair, is primarily how you should use an encyclopedia.) The degree to which you should trust it as your sole starting point for research in an unfamiliar area is anti-correlated with the length of the article's Talk page.
I used Abbyy Finereader for several years. I loved it. I completed some large projects with it. Modern VLMs put classic FineReader to shame for processing low-resolution/degraded/non-standard text.
I'm personally using the small Qwen 3.5 models. If you have an OCR problem, Mistral OCR 4 is probably great. Open weights models that you can run on a laptop may also work great.
If the Swiss thought it was in the national interest to exit the Nuclear Non-Proliferation Treaty and crash-develop a nuclear deterrent, I think that they could achieve nuclear breakout quickly.
"In an all-out existential battle" involving nuclear weapons, the United States won't be affected by the presence or absence of domestic car factories either. World War II could soak up years of total warfare effort from the belligerents, and still have factories and governments intact to send more soldiers and bombs toward the enemy. I don't think that can happen now that countries as poor as North Korea can make nuclear weapons.
The newer drones may have dedicated warhead designs now, but the concept is similar. They carry high explosive warheads that aren't sold to civilians. It's the same reason the militias don't already have anti-tank weapons.
The raw training data is so large that very few parties could host it for free even if there weren't copyright barriers.
But I think you could have a full open source training software pipeline that's set up to work with Wikipedia, Common Crawl, Books3, Library Genesis, Anna's Archive, and whatever other useful data sets people can name. There would just be a step where you have to provide your own copy of Library Genesis (or whatever subset of it you have managed to obtain).
Note also that Anthropic's definition of "unsafe" encompasses "competing with Anthropic."
In light of the ability of recent models to accelerate their own development, we’ve implemented new interventions that limit Claude’s effectiveness for requests targeting frontier LLM development (for example, on building pretraining pipelines, distributed training infrastructure, or ML accelerator design). Using Claude to develop competing models already violates our Terms of Service, but enforcing this restriction through our safeguards avoids accelerating the actors most willing to violate these terms.
Unlike our interventions for cybersecurity, biology and chemistry, and distillation attempts, these safeguards will not be visible to the user. Fable 5 will not fall back to a different model. Instead, the safeguards will limit effectiveness through methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT). These interventions will not affect the vast majority of coding work. We estimate they will impact ~0.03% of traffic, concentrated in fewer than 0.1% of organizations. When these interventions are active, we expect them to have minimal behavioral impact on the model except to limit its effectiveness in developing frontier LLMs. Claude will still respond helpfully to user requests. We’ll continue to improve the precision of our detection methods following the launch of this model.
(From the model card document)
I didn't previously understand that they interpreted "Using Claude to develop competing models" so broadly. I thought that meant something like "our ToS disallow distilling our models."
Too bad. I'll continue to use Claude for now, because it's quite effective, but in the long term I don't want powerful models like these to be controlled by any one nation or company.
I remember when I had to wait minutes to get a high resolution image over a dialup connection. When computer and communications hardware advanced enough that I could get 30 high resolution images every second, there were brand new uses. In the case of LLMs, I could imagine that much faster operations allow you to introduce them as parts of systems that need to react to the real world at high speed, like factory equipment. Showing that a model can do the usual LLM tasks at extremely high speed is just a demo proving that the approach works.
From the description I thought that a degraded capacitor or lack of lubrication made the blower not start on its own, but the blower (and the whole furnace) would work if given a manual startup spin by hand.
More good news from Ember, according to their Global Electricity
Review 2026 [1]:
Solar power increased by a record 636 TWh to reach 2,778 TWh in 2025, a 30% increase from 2024.
Wind saw the second-largest increase, growing 205 TWh (+8.2%)
Driven by record solar growth, low-carbon power generation increased by 887 TWh in 2025, outpacing electricity demand growth of 849 TWh. Solar power alone met 75% of the net increase in electricity demand. Together with wind, the two sources met almost all (99%) demand growth.
For the first time in 100 years, renewables (33.8%, 10,730 TWh) overtook coal power (33.0%, 10,476 TWh) in the global electricity mix as continued rapid growth in solar and wind pushed the share of renewables above a third of global generation. Coal power dropped 63 TWh (-0.6%) in 2025, marking the first fall since the Covid-19 pandemic in 2020. Combined with continued electricity demand growth, this meant coal fell below a third of global generation for the first time in history.
For comparison, I have collated information from the International Atomic Energy Agency's Power Reactor Information System. The fastest that nuclear power generation ever grew was 213 TWh added in 1985. Since the year 2000, the fastest growth year was 2004, with 111 TWh added.