The Y axis is kilograms, the X axis days. My Fast 800 weight loss journey. Most of the health stats (BP/pulse and others) I've monitored are largely unchanged but I feel much lighter and quicker than at the start.
This one was a tricky 4.
Wordle 1,184 4/6*
⬛🟩🟨⬛⬛
⬛🟩⬛🟨⬛
⬛🟩⬛⬛🟩
🟩🟩🟩🟩🟩
The Soviet InterNyet
https://aeon.co/essays/how-the-soviets-invented-the-internet-and-why-it-didnt-work
Soviet automation, data & automation policy/failures, area lesson in how not to do things - here in the west, in a similar time frame, it was laundered through, and in collaboration with, private industry/capital and individuals. Inria and Fraunhofer/Max Plack being good European examples of what was happening in the US*. The UK has always been far more university-centric/laissez faire/less dynamic.
* The German policy and British attitudes have 18th/19th century roots and the US did it better. I'm woefully ignorant about Chinese policies. I do know it's not government or private that wins, it's both, a least in Europe/the US, although there are arguments about the balance.
Haha, just got recommended this via Instagram/Facebook's advertising algorithm(s). Which is tangentially related to my last minor rand if you've read Spufford's Red Plenty which is a very good book for technologists of all stripes, although the pacing in the first half is better IIRC.
https://www.faber.co.uk/product/francis-spufford-in-conversation/
Begun, the open source AI wars have
https://www.theregister.com/2024/09/14/opinion_column_osi/
Training data is the biggest unexploded litigation timebomb right now. Although it will probably be money suing money rather than the overwhelming majority of users whose data was scraped up bar class action. The audit trail on weighting is non-existent. The maths is already free and currently big money is benefitting from it, although the customer benefits are presently debatable. By that I mean: take the AI generated picture accompanying the article - the bulk of the work involved with the article is writing the article itself. AI can copy and collate other original sources but not from primary sources, via phone and email, yet, it'll get there and go wrong a lot, and not without a great deal of prompting and correction. In asymptotic terms it's not the big productivity factor. Nice but not great. Maybe in image editing and a number of other areas it's extremely useful and a time saver. That's not a huge chunk of the world's productivity, which is largely dealing with complicated systems where AI is not a magic bullet. Yet. It could be. LLMs are the start and machine learning, in a more general sense, has been here for decades and is/has been extremely important already. Automation goes way beyond AI.
An excellent video. The late 1950s/early 1960s was really the genesis of so much now cf. J.C.R Licklider's Man Computer Symbiosis (1960), Robert Bellman (1950s), Computers and Thought by Feigenbaum & Feldman (1963) - a compendium of the state of the art then, and stuff like the perceptron. I feel like we've reached the point where there is enough available computing for ideas that already existed to be implemented without finesse and many of the people taking credit, while totally deserving of praise because doing stuff is hard, shouldn't be viewed in isolation. See also Hypertext and SRI/Douglas Engelbert (PARC and others). Giants on the shoulders of giants etc.
Somewhat GIGO in, somewhat GIGO out*.
Have we stopped to think about what LLMs actually model?
https://www.theregister.com/2024/08/30/ai_language_cognition_research/?td=rt-3a
* I'm not a machine learning sceptic. Increases in processing power will change the world much as dynamic programming did behind the scenes. I am sceptical about some key framing - not least intelligence, which seems like very anthropomorphic concept divorced from the parameters imposed by an environment or even shifting environments. FLOPS are underpants and people consuming the hype are gnomes.
PSNR worked best objectively but SSIM correlated most with my subjective judgments. Presumably PSNR HVS would be similar, which I didn't have time to tabulate. Usual Saturday.
Large chunks of the internet could easily be replaced with generative AI because people are just repeating, at best, conventional wisdom and at worst the marketing for various products or standards.
So
I spent the last 2 hours working out PSNR values from a reference with various parameters. And the people indistinguishable from generative AI are mostly correct, although they don't know why, or by how much or under what circumstances.
Not that it's that important. It just annoyed me.
A brave new world of computer assistance. 