cedricb_42

@geoffreyirving We tried that! The vintage models can just barely start to do simple things with Python, purely from in-context learning:

Announcing Talkie: a new, open-weight historical LLM! We trained and finetuned a 13B model on a newly-curated dataset of only pre-1930 data. Try it below! with @AlecRad and @status_effects 🧵

You should definitely not use such evil sites that steal from the Good and Honorable publishing companies just to make research available to the tax payer who paid for it.

What exactly is the socialist case for stealing from the Louvre It’s a publicly owned museum, created when the French revolutionary government seized the royal palace from the monarchy and opened it to the public Hasan Piker just equates leftism with being anti-social

@LAsixx9 @BeefcakeDillon

@BeefcakeDillon Love is a lie made up to keep people tied down and content to a shitty life and compell them to raise kids. Single people are happy and free but society creates the narrative that we're the bad guys

The thing about casual sex is it has little to no appeal to a real lover. The thought of having sex with someone I do not care deeply for feels so hollow and meaningless

there is no free lunch... risk is compensated with a risk premium... the 17.5% are hardly guaranteed...

@mayukh_panja That's not a fair comparison tho... You picked a bad day in Berlin and a very good day in Mumbai....

lol. this is going to really excite the “biggest bubble to burst is quality of life in abroad” guy.

@mayukh_panja @snr_boost @rahulj51 Yeah, it's super cheap

@snr_boost @rahulj51 for a quick trim?

I dread going to the barbers here in Berlin. 90 % of the time they are unfriendly and when they are not unfriendly they are in a hurry, and its almost like they are annoyed that they have to do any work. There was a new guy today and he decided to shave my neck (after my trim) with a bare blade without any cream, or even water?? What the fuck? Obviously little beads of blood appeared and I pointed it out, but he just shrugged and kept scratching away with the blade on top of the cuts? Who the fuck does that? He seemed the least bothered by it all and just moved on to chatting away with a friend after he was done. And there are always like 5 random guys lounging around in every barber store. I know the solution is to pay more and book a fancier one that shaves you for 45 minutes. But it shouldn’t be so hard to find someone that is just polite and pays some attention to what they are doing. I wonder if this is universal or it is just me who is just unlucky.

@RamboVanHalen Same in Germany

Today I learned that catch and release fishing is illegal in Switzerland. It's an animal rights thing. They feel it's cruel to subject a fish to being caught more than once. It's still legal to kill the fish, but not put it back in the water alive. If questioned by the police, conservation oriented anglers have to lie and say that the fish "slipped out of their hands".

You can have the a SOTA model just duplicating some layers. The author discovered that today's big AI language models (like those with 70+ billion parameters) have a kind of internal "brain structure." Early layers handle understanding input text, late layers prepare the answer, and the middle layers do the actual deep thinking/reasoning. He found that if you carefully duplicate (repeat) a specific block of those middle "thinking" layers, without changing any numbers in the model at all the AI gets noticeably smarter on hard benchmarks. It's like letting the model think twice through the same reasoning steps. He called this technique RYS ("Repeat Yourself Self") and applied it to a strong 72-billion-parameter model (a fine-tuned Qwen2-72B called calme-2.1). He repeated 7 specific middle layers (layers 45–51), turning the 72B model into an effectively ~78B model. No training, no fine-tuning, no merging just clever re-use of existing layers during inference. This single change pushed his model (RYS-XLarge) to #1 on the Hugging Face Open LLM Leaderboard at the time. Key improvements (compared to the original model): +17.7% on MuSR (multi-step reasoning) +8.2% on hard MATH problems Smaller but consistent gains on BBH, GPQA, MMLU-PRO Only a small drop on one benchmark (IFEval) He found this sweet spot by running lots of automated tests ("brain scans") on just two gaming GPUs (RTX 4090s), using clever proxy tasks (hard math questions + emotional reasoning scores) instead of running the full expensive leaderboard every time. Why it works? The middle layers form indivisible "reasoning circuits" that the model developed during original training. Copying the whole circuit once more gives the model extra thinking time/power without breaking anything similar to re-reading a difficult textbook chapter to understand it better.

@jandr1810 Google search Trends werden mit extra noise versehen, daher sind solche kleinen Spikes nicht aussagekräftig @FlorianGallwitz hat da einiges drüber getweeted

warum wurde mein full government name am 19.02. 100 Mal (!!!) in israel gesucht??? und mal was ganz anderes wie erkennt man eigentlich palantir drohnen

the honest reality is that organic social interaction has basically collapsed for a lot of society, especially in cities where paradoxically everyone is surrounded by ppl but nobody talks to each other. 50 years ago, you couldn't have predicted this at all.

@paleonormie @meteoblue does something somewhat similar to this. they show the confidence levels of the prognosis and also allow you to manually see distribution of forecasts made by different models. I can fully recommend them!

weather apps should have an advanced feature for smart people where they publish a range of possible high temperatures with error bars instead of one middle estimate that will always be wrong

"Field marshal Haig has formulated a brilliant new tactical plan to ensure final victory in the field" "Would this brilliant plan involve us launching drones at Dubai very slowly?"

me after reading the anthropic blog post

@theo @ji1kyo You look like yann le cun 😆

@boyteased @ji1kyo I gotchu

he sent these knowing we’ll never actually get him in glasses again beautiful how could you

@vistaaura222 stunning. unfortunately it only looks like that a few weeks of the year, the remaining time it looks grey and depressing...

Pentland, Scotland

In 2016 I was working at a company that wasn't really sure what they wanted to do. I was there working on an Android app that was destined to fail. The founders were interested in making a payment app of some sort. They told our intern to: 1. Make a list of all the payment apps in Hong Kong. 2. Download and try out all these apps. 3. Write a report about each of them and present it to us. Pretty reasonable. Most of the apps he tried out didn't really work, or didn't seem to have P2P payment function. Finally he tried one that seemed to work. Intern: Angus I think this app works, but I can't really verify it. Angus: I dunno, let's make accounts and try sending money to each other. Intern: Ok. I stopped coding the Android app I was working on. The intern was watching me sign up, and to be a cool guy I signed up with the username "Bonerlord". The intern found this very funny and so he signed up as well, with the username "Viagramaster". What a pair of funny guys. We poked around the website, it had features like: 1. Search for friends 2. Add friends 3. Send money The search for friend's feature had really aggressive autocomplete. If you typed in the letter 'A' it would show you a list of all users starting with the letter A. So I did that, and clicked "Add friend on like 100 users". Just to test out the app. We did that for a bit. Couldn't figure out how to add money to our accounts and then stopped using the app. About an hour later I get a call from an unknown number. I hung up on it. The intern gets a call soon after. He answers it. He talks for a bit and sounds really worried. I look at him. He covers his phone and says "it's that app". I take the phone from him. The guy on the other end is really angry. Phone man: Who are you guys? Why are you hacking my app? Me: We didn't hack your app. We were testing it out. We did everything through the webpage. Phone man: You spammed all our customers about viagra. I realised everytime we clicked "Add Friend" it sent an email to that user with something like "Bonerlord has sent you a friend request". Me: No no. My username is Bonerlord, I just added random users on your website. I typed in the letter 'A' and it showed me everyone who had a username starting with A. I just clicked a bunch of people to test your app. Phone man: Hmmm okay. Me: Anyway, we're building an app maybe we can integrate with your system. Can we set up a meeting. So we got his contact and then met with him in real life. The meeting went nowhere, but when me and the intern arrived he said "Ah this must be Bonerlord and the Viagramaster".

@itsrapha83 The twitter algo pushed one of your posts into my timeline and I immediately loved the shots, so I checked out your profile and immediately clicked follow

How did you find my account? Do you remember what post brought you here?

During my first year as a PhD at ETH (it's special here that we need to supervise master students for projects), I observe this: Many master students devote 100% time to courses (teaching all those classic knowledge) and do not know about the exciting research frontiers at all.