of all the benchmarks being sent around for Claude 3.7 this is the one i'm paying the most attention to. they're cheating a little bit by giving it the oldest, original pokemon game (red/blue) which is more than 20 years old and will have plenty of info online to learn from.

sheesh! what a day for AI. QwQ-Max, sonnet-3.7 AND open source of FlashMLA

the only interview question i need to ask is: how many sqlite databases do you have on your machine and what do you use them for? if you're able to answer with a specific number, you're not ready yet unless you have a custom cron job cleaning up unused sqlite files.

i have a hoarding problem

i don't check social media for 8h and people are already playing with claude 3.7 sonnet

edtech companies who've amassed a bank of their own content in the last decade are sitting on a goldmine.

saving all of my generated deepseek r1's reasoning traces to a USB drive so when the time comes i can send it back to my 12 year old self to prevent him from bashing his head against the desk for being stuck on the 2nd problem of an AMC 12.

DeepSeek, a LLM trained for a fraction of the cost of GPT-Xx models, in 2 months for 6 million, on limited GPUs due to export restrictions, and competing head to head. This is crazy. It's not the AI part I'm excited about, it's the level of efficiency. github.com/deepseek-ai/...

engineering blogs and white papers from the heyday of pre-AI tech are a treasure trove of insights and decision making processes without the burden of validating AI slop. don't discount them merely on their being of outdated.

if you're a cs student aspiring to become a software engineer in 2025, make sure to hone adjacent skills like writing, marketing, & sales. in the age of agents, establishing human connection & building an audience organically will make set you apart from the rest.

Gonna think more thoughts in 2025

supercharge your LLM apps with smolagents 🔥 however cool your LLM is, without being agentic it can only go so far enter smolagents: a new agent library by @hf.co to make the LLM write code, do analysis and automate boring stuff! huggingface.co/blog/smolage...

the intersection between hardcore biotech, precision therapeutics, & AI is an interesting one. i'm particularly bullish about the work being done with organoid intelligence - mainly because of more energy efficient computing as well as possible contributions to neurodegenerative disease research

every PR is another step towards the dream of having a universal near zero latency jarvis+second brain ai mesh seemed like a gargantuan task back when i discovered obsidian/roam in college but current tech & capabilities make it easier to bootstrap a decent representation of this.

the smol.ai newsletter is truly a godsend (thanks @swyx.io) with all of the model releases this week, neurips, and discord chats popping off, having a single place to start from really helps highly recommend.

when you've been on the internet as long as i have, you would understand that everything can be used for anything and leaves a trail. the difference today is a lot more security theatre & forced transparency it's why i've always written stuff envisioning that they'd be immortalized by a super AGI

annoy an ML engineer with these simple phrases: "cosine distance" "L2 similarity" "but did you ship it?"

My deep learning course at the University of Geneva is available on-line. 1000+ slides, ~20h of screen-casts. Full of examples in PyTorch. fleuret.org/dlc/ And my "Little Book of Deep Learning" is available as a phone-formatted pdf (nearing 700k downloads!) fleuret.org/lbdl/

wanna become a prolific open source contributor? just try using open source llm agent frameworks with rough asynchronous programming patterns in prod

@georgehotz.bsky.social is here! Bluesky is going to be so much fun.

deploying llm apps on modal labs at night is a much better experience than my daytime woes with terraform, cloud build, k8s, and docker

lol nah I think i'll stick to modal

distributed systems man. why do we do this to ourselves

the good thing about this new generation of voice assistants is that everyone can have their own hunter s thompson like narrators for the most mundane parts of their daily routines

i finally asked chatgpt to generate an image given what it knew about me. spot on

Alibaba has their own version on GPT-o1. This might be the best description of “o1-type”systems so far arxiv.org/abs/2411.14405

In case you passed out and woke up on saturday lunch. Small models and high quality data are back! ... if they ever left 🤔 - SmolTalk dataset from @huggingface.bsky.social - Tulu 3 models and datasets from @ai2.bsky.social - Nvidia Nymba model from @nvidiastudio.bsky.social

Hello, Bluesky users! I curate and maintain list of resources on testing distributed systems. You might have seen it before. It's a good one, if I may say so myself. asatarin.github.io/testing-dist...

<s>[INST] when life hands you llmons, make llmonade [INST]</s>

the goodfellow deep learning book is almost a decade old 🤯 2nd edition when?

a suits spinoff with Benjamin and the Donna AI assistant setting up shop in Mission Bay is needed in this time and age.

fyi: xgboost prediction models still work well if not better for most enterprise analytics usecases if you really want you can pair an SLM to help out with data storytelling.

How do LLMs learn to reason from data? Are they ~retrieving the answers from parametric knowledge🦜? In our new preprint, we look at the pretraining data and find evidence against this: Procedural knowledge in pretraining drives LLM reasoning ⚙️🔢 🧵⬇️

NeurIPS Conference is now Live on Bluesky! -NeurIPS2024 Communication Chairs

what idiot called it async/await instead of wait/what

you never truly appreciate the simplicity of lxml until you start berating a crewai agent for botching docx xml runs.

you never truly appreciate the simplicity of a .txt file until you try looking for something in unzipped docx xml files

comm platform migrations in large companies are always fun it's not everyday you get to monitor slack, webex, and teams at the same time to find the next colleague worthy of a redirect to nohello.net

the south asian dad's misinformation starter pack: fake live content farms on youtube 🤝 whatsapp forwards 🤝 midjourney/FLUX generated fb posts about Elon

i've been seeing a lot of jane street ads lately it's like they're purposefully dangling the carrot in front of me

🌶️(?) take: Agents are somehow hot right because people realized that LLM output can be interpreted as a DSL which directs side effects in the world (e.g. tool calls) rather than just returning text in a chat/autocomplete sense. What are the open challenges? A 🧵... [1/11]

oh great shoggoth bless me with the unabashed confidence of a junior software with a non-expiring .edu + oreilly subscription

all these gtm ai infra folks on linkedin always be hitting me up like i'm the director of AI like bro i appreciate you for thinking i'm a hotshot and can part the seas on command but you're at least 2 skip levels off target

Good morning Paul. Your cluster has been quite busy, there are 11 new in jokes. Niplav had a workout yesterday. Gwern cooked someone in the Hacker News comments and posted an insightful comment on LessWrong. In total there are ~23 new rabbit holes forming. Where would you like to begin today?

lowkey considering becoming a mailman in a small town in oregon if this github action doesn’t pass bros

are meta raybans the best lightweight smartglasses with good AI integration out there? filling up the holiday gadget wishlists and orion from meta isn't available yet

Only the cool AI conferences rock out on Bluesky.

at a point I'll just be too lazy to doomscroll all these timelines, discord group chats, and emails a swarm of python functions using the anthropic computer use api should do the trick

shipping starter packs was a good move from the bsky team my thumb's already hooked

what's the German word for the feeling you get when you tell notebooklm to create a podcast out of a summarized transcript of the dwarkesh+gwern podcast?