ThatOtherThunderGod🐇

No, @tylenol didn’t actually mean that its products aren’t recommended during pregnancy. Unravelling a poorly-worded 2017 tweet and other misinformation about acetaminophen and pregnancy that's been spreading this past week. #misinfo #healthnews pressprogress.ca/no-tylenol-did…

#NEW: In Jan @VancouverPD crafted a narrative about how antisemitism and the "Israel-Hamas war" were the biggest cause of spike in hate crimes. Now, FOI data reveals other minority groups reported higher increases in hate crimes: Lates for @TheTyee tinyurl.com/3m39xdk4

StarFish expands to California! We are thrilled to welcome Omnica to the StarFish family. This acquisition expands our reach to California, enhancing our capabilities and strengthening our commitment to serving MedTech innovation across North America. starfishmedical.com/news/starfish-…

The civilian body that reviews complaints against RCMP has assessed multiple cases & found the C-IRG unit at #FairyCreek "repeatedly acted in a way that is contrary...to the rule of law" This is in a report about one man's case. My latest for @NatObserver tinyurl.com/3rtx6pcv

The insane irony of this is that each one of these failed multiplications requires performing trillions of successful multiplications

"NNs just want to learn" no. they just want to cheat. they are wily devils that will overfit to any spurious correlation present in your data that you didn't even know was there. they are demons

Everything wrong about our perception of machine learning and AI is due to excessive anthropomorphism -- which the field actively promotes.

@kalomaze There's been some (only pre-print, but still actually public) work in this space, focused on the <1B param regime, for obvious reasons. They found depth > width for "reasoning" tasks (more attention head interactions?), but had no comparison for "memory" arxiv.org/abs/2402.14905

there seems to be a correlation between intermediate dim size and better compression of highly specific/niche knowledge unfortunately, shapes during pretraining are basically an afterthought, barely anyone studies how these ratios change the model (even tho they do)

i find it peculiar that Gemma 27b has a ratio of 90% MLP / 10% Attention for each layer, rather than 80/20 which seems to be common (Mistral, Llama3, etc) it also has a much larger MLP expansion ratio - 8x instead of 2.3x (seen in Mistral 123b) - prob why it's good with knowledge

I Wake Up

Making good software would be pretty innovative

NEW: In my first for @NatObserver, I tried to find an answer for how much water #cryptocurrency it takes to mine cryptocurrency in Canada and whether there are crypto-specific water restrictions during droughts. TLDR: a lot more than we know, and nope. nationalobserver.com/2024/07/2/news…

Got an amazing tattoo with artwork from @gabrielledrolet Thanks so much for helping me immortalize Hopscotch!

April 8th, 2024, a Threat Actor operating under the moniker "USDoD" placed a large database up for sale on Breached titled: "National Public Data". They claimed it contained 2,900,000,000 records on United States citizens. They put the data up for sale for $3,500,000. National Public Data is a background check and person lookup and verification company located in Coral Springs, Florida. The company offers API data lookups to other companies. In summary – they're a lowkey databroker. Last week we were informed USDoD intends on leaking the database. We requested a copy in advance to confirm the validity of the data. We reviewed the massive file – 277.1GB uncompressed, and can confirm the data present in it is real and accurate. We searched up several individuals who consented to having their information looked up. 1. The database DOES NOT contain information from individuals who use data opt-out services. Every person who used some sort of data opt-out service was not present. 2. People who did not use data opt-out services and resided in the United States were immediately found. It showed their: - First name - Last name - Address - Address history (3 decades+) - Social security number It also allowed us to find their parents, and nearest siblings. We were able to identify someones parents, deceased relatives, Uncles, Aunts, and Cousins. Additionally, we can confirm this database also contains informed on individuals who are deceased. Some individuals located had been deceased for nearly 2 decades.

Watching tech companies pour billions into trying to create the Torment Nexus, but the tech is so unreliable they only manage to release the Mild Annoyance Nexus

@gabrielledrolet 250g!

doing a giveaway to celebrate 26k. first person to guess pidgey’s weight gets a print of their choice in the mail. only one guess per person. guesses have to be in grams. woman’s hand for scale

"It is impossible to tell the difference between our gold-painted lead and actual gold, therefore we have really transmuted lead into gold" You shouldn't ask random people (with only a faint idea what gold looks like) if your shiny object is gold or not. You should ask experts who have a proper definition of gold and who know how to perform an acid test.

This is all in a refugee camp, just so everyone knows.

Israeli concentration camps, confirmed by whistleblowers working at one of the sites. Israel is committing — and the U.S. is funding — the second Holocaust.

KBJ: what if the city decided to execute homeless people? Gorsuch saves lawyer: that would be both cruel and unusual, right? Lawyer: hemming/hawing... Gorsuch: how about just a yes on that? Lawyer: yes

The invasion of so many of LLM-based products has sparked furious debate about their reliability, particularly in high-risk applications. Thor discusses the risks of using AI in medical devices and provides recommendations on how to use AI responsibly. starfishmedical.com/blog/using-ai-…