The Lever: "The AI Boom Just Jacked The Price Of Your Electronics.. Prominent computer chip part maker Micron is abandoning its consumer-side business in favor of artificial intelligence, enabling its competitors to raise prices."
The Lever: "Good things are happening! One of the most restrictive labor laws in the country is overturned, a predatory rideshare billing scheme could be nixed, lawmakers take on personalized pricing, and a state gets serious about bottled water."
Firstpost: "Brazil's Lula warns armed intervention in Venezuela would be 'humanitarian catastrophe'"
Deir ez-Zor, Raqqa, Palmyra all seem to be outside known ISIS areas in Syria... But I guess it is a message to someone, somewhere.
"U.S. launches strikes in Syria targeting Islamic State.. A U.S. official described it as "a large-scale" strike that hit 70 targets in areas across central Syria that had IS infrastructure and weapons... Deir ez-Zor and Raqqa provinces and in the Jabal al-Amour area near the historic city of Palmyra"
Why does GPU restrictions effect Chinese LLM ("AI") efforts? A lot of training approaches in machine learning divide the training data into batches and multiple CPUs and GPUs work on them. Many ML algos are said to be "embarassingly parallel", ie they can be parallelized at ease.
Then one could think if China has access to chips half as good due to export restrictions, can't they double the number of chips and get equal performance? It seems one major hurdle is in the sync phase of said LLM training algos. To synch they need to communicate, and with increasing chips that communication bottleneck grows non-linearly for LLM training. NVidia has an optimized interconnect tech for its GPUs, Google developed its own for its TPUs, China is behind in that regard (5-10 yrs depending on who you ask). That means China not only needs to develop faster chips but also faster interconnect methods.
Is this a big loss? I don't think so. I don't see LLMs as a major strategic advantage in any field. If you need focused neural net applications you can train those stupid black boxen with less hardware, no need for thousands of GPUs and their interconnect.
BTW the SOTA for recommendation algorithms are not neural net based. Let that sink in. The machinery that powers "intelligent" LLMs is not the top choice for a major area in machine learning.
"@brohrer@recsys.social
Apple TV just recommended me an M Night Shamalan and I'm trying not to feel insulted"
#MS #LLM
![[-]](https://media.zeroes.ca/media_attachments/files/115/754/953/903/347/693/original/7d9da9790bc7c1d6.png){kind=link}
Except Hinton probably, but he's gone senile... Sutskever is his student, we can defer to him.
He argues research needs to focus on new AI approaches that learn how to learn. Current LLMs are too finished, too polished in certain tasks, and they still suck at many tasks human learn easily (like learning chess, learning how to drive). A lot of AI heavyweights are advocating going back to the drawing board these days.
#Sutskever
"The future of intelligence | Demis Hassabis (Co-founder and CEO of DeepMind)"
"@andrewdessler@mastodon.world
Zeke Hausfather makes his predictions for 2026 and 2027 temperatures. I don’t need any fancy science to tell you that both years gonna be en fuego"
![[-]](https://s3.eu-central-2.wasabisys.com/mastodonworld/cache/preview_cards/images/074/944/137/original/b5ef180776233694.jpg){kind=link}
Chapelle: "Don’t forget what I just went through... Two years ago, I almost got canceled right here in the United States for transgender jokes. But I gotta tell you something: Transgender jokes went over very well in Saudi Arabia"
Fucka me, fucka you. UA attacks RU ships, they pay back in kind. UA has probably more to lose in this exchange than Russia.
"A Russian drone hit the civilian vessel VIVA, which was transporting sunflower oil through the Black Sea grain corridor"
"With Attacks on Oil Tankers, Ukraine Takes Aim at Russia’s War Financing"
This is good.. The government needs make businesses its bitch. For too long government has been private sector's bitch. That needs to change.
NPR: "President Trump said the administration has reached agreements with nine more drugmakers to bring their U.S. drug prices more in line with what other wealthy countries pay. Fourteen companies in total have now reached what the administration calls most-favored-nation pricing deals. The companies that took part in Friday's announcement were: Amgen, Boehringer Ingelheim, Bristol Myers Squibb, Genentech, Gilead Sciences, GSK, Merck, Novartis and Sanofi.
Top federal health officials, President Trump and Pfizer CEO Albert Bourla meet in the Oval Office to announce an initiative on drug prices. They agreed to charge the U.S. government no more for new drugs than the prices paid by other well-off countries. The agreements will allow state Medicaid programs to access lower prices from the nine companies"
Reference
Nations and Nationalism, Culture, Narratives
Rome, The First Wave, Religion