burnytech.bsky.social - Profile | ThreadSky | a Reddit-style client for Bluesky

THIS IS HUGE! Researchers at Memorial Sloan Kettering Cancer Center say results of a Phase I trial of an mRNA vaccine aimed at preventing pancreatic cancer show promise. The vaccine PREVENTED the cancer from coming back in patients who had already received treatment for MORE THAN three years. 🧪🧵⬇️

submitted 2 days ago • 344 comments

oops

submitted 2 days ago • 2 comments

I love the new citations preview in overleaf, this is really useful! 😍 @overleaf.com

submitted 1 day ago • 3 comments

They can jump now

submitted 2 days ago • 5 comments

How can AI help physicists search for new particles? The ATLAS and CMS collaborations are using state-of-the-art machine learning techniques to search for exotic-looking collisions that could indicate new physics

submitted 3 days ago • 1 comment

submitted 3 days ago • 6 comments

I have to admit that I've been surprised at the level of hate at AI LLMs. There is finally a tool that allows those without access to an advanced education to self learn, anywhere , anytime. To get answers to questions, (even if sometimes wrong) , that they would otherwise never have access to

submitted 6 days ago • 2611 comments

Agree with this. LLMs have all sorts of abuse cases, but also have real use cases. It’s a shame to miss out on the latter because of anger at the former. I know I’ve found them helpful in learning things and getting a feeling for a field. Just gotta keep your wits about you.

submitted 6 days ago • 24 comments

Is your model ready for the real world? Learn how to evaluate the performance of deployed image classification models and gather valuable user feedback in Part 3 of David Martin's series.

submitted 7 days ago • 0 comments

Why is life on Earth left-handed? No-one knows, but the answer could be found in comets, asteroids and on Mars www.skyatnightmagazine.com/space-scienc... #astrobiology #astrochemistry #biochemistry

submitted 7 days ago • 1 comment

Simply masking 15% of input tokens + next-token prediction (NTP) can significantly boost LLMs on key information retrieval & long-context reasoning—without extra compute! MEAP (Mask-Enhanced Autoregressive Prediction)

submitted 7 days ago • 2 comments

@alphaxiv.org used Gemini 2 Flash to build Cursor for arXiv papers Highlight any section of a paper to ask questions and “@” other papers to quickly add to context and compare results, benchmarks, etc.

submitted 17 days ago • 1 comment

Professors waste ChatGPT's potential. It's not a search engine, it's your co-pilot: • Write megaprompts for research design • Build custom GPTs for repetitive tasks • Get unbiased reviews of your drafts Stop using a supercomputer to fetch some papers.

submitted 16 days ago • 0 comments

Consistent Jailbreaking Method in o1, o3, and 4o Discussion

submitted 16 days ago • 0 comments

Fun to watch Deep Research basically go through the same learning journey I did in my first year PhD classes on organizational theory in order to answer a question, though this was in a few seconds, rather than a semester (and I hope I grappled with the concepts more deeply myself?)

submitted 17 days ago • 2 comments

Dynamically scaled softplus attention Builds on log-scaled attention, replacing the softmax exp with softplus but keeping the normalization. Reminiscent of sigmoid attention - which it reports as also outperforming softmax results when applying the same modifications. arxiv.org/abs/2501.13428

submitted 16 days ago • 1 comment

LMArena's Arena-Price Plot! 💰📊 An interactive plot of price vs. performance trade-offs for LLMs. lmarena.ai?price

submitted 16 days ago • 1 comment

GRPO with Verifiable (Binary) Rewards Is an Adaptive Weighted Contrastive Loss by Youssef Mroueh The goal of this short blog is to understand GRPO that was used successfully to train Deepseek models. We will limit our analysis to binary rewards or what Tulu authors calls RLVR.

submitted 20 days ago • 1 comment

Making LLMs run efficiently can feel scary, but scaling isn’t magic, it’s math! We wanted to demystify the “systems view” of LLMs and wrote a little textbook called “How To Scale Your Model” which we’re releasing today. 1/n

submitted 19 days ago • 3 comments

How to Scale Your Model This book aims to demystify the science of scaling language models on TPUs: how TPUs work and how they communicate with each other, how LLMs run on real hardware, and how to parallelize your models during training and inference so they run efficiently at massive scale.

submitted 19 days ago • 1 comment

open-Deep-Research by huggingface as posted by @aymeric-roucher.bsky.social An entirely open agent that can: navigate the web autonomously, scroll and search through pages, download and manipulate files, run calculation on data...

submitted 19 days ago • 1 comment

I got an email reminding me to "increase the visibility of my work". RL did some press on X but I haven't posted on my new home about this. And so.... Our first-gen ASIC/FPGA decoder paper was published, in .... drum roll, please ......... Nature Electronics! www.nature.com/articles/s41...

submitted 20 days ago • 2 comments

What to know about DeepSeek youtu.be/0eMzc-WnBfQ?... In which we attempt to figure out MoE, o1, scaling, tech reporting, modern semiconductors, microeconomics, and international geopolitics.

submitted 19 days ago • 1 comment

Revisiting the Multi-planetary System of the Nearby star HD 20794: Confirmation of a Low-mass Planet in the Habitable Zone of a Nearby G-dwarf astrobiology.com/2025/02/revi... #astrobiology #exoplanet

submitted 19 days ago • 0 comments

⭐ The first foundational model available on @LeRobotHF ⭐ Pi0 is the most advanced Vision Language Action model. It takes natural language commands as input and directly output autonomous behavior. It was trained by @physical_int and ported to pytorch by @m_olbap 👇🧵

submitted 19 days ago • 5 comments

n8loom A library for generating trees-of-thought - the kind used in MCTS, variants of majority-voting, etc. - efficiently by splitting the kvcache into fragments at each node, and dynamically concatenating the results together when generating.

submitted 20 days ago • 2 comments

Language Models Use Trigonometry to Do Addition They discover numbers are represented in these LLMs as a generalized helix, which is strongly causally implicated for the tasks of addition and subtraction, and is also causally relevant for integer division, multiplication, and modular arithmetic.

submitted 20 days ago • 5 comments

New (short) paper showing how the in-context inductive biases of vision-language models — the way that they generalize concepts learned in context — depend on the modality and phrasing! arxiv.org/abs/2502.01530 Quick summary: 1/5

submitted 19 days ago • 1 comment

ReAG - Reasoning Augmented Generation - No chunking, splitting vectorizing bs - Stateless, no vector DBs etc. - Supports any model (deepseek, o3-mini et al) - Reasoning traces - Metadata filtering - Typescript, Python support

submitted 20 days ago • 2 comments

I followed tons of STEM people to see STEM on my feed, but my feed is full of politics from people I don't follow. Is there a way to fix this? Marking 1000 posts as "show me this more" and "show me this less" doesn't seem to work. :(

submitted 19 days ago • 3 comments

Not good. SiliconFlow, a Chinese LLM provider, is offering DeepSeek R1 inference services on the Huawei Ascend 910B. This is what I am afraid of: competition from China due to our misguided ban. So far, none of the U.S.-based LLM ASIC providers offer similar services: - Groq - Cerebras - SambaNova

submitted 22 days ago • 3 comments

What happens if you put an AI in charge of national defense? In war games, LLMs tend to escalate & do arms races. Base models are more aggressive & unpredictable. The authors speculate that it is because there is lots of training data on escalation, little on de-escalation. arxiv.org/pdf/2401.03408

submitted 22 days ago • 8 comments

It output about 2 tokens per second. Painfully slow. How To Run Deepseek R1 671b Fully Locally On a $2000 EPYC Server digitalspaceport.com/how-to-run-d...

submitted 22 days ago • 1 comment

Since everyone wants to learn RL for language models now post DeepSeek, reminder that I've been working on this book quietly in the background for months. Policy gradient chapter is coming together. Plugging away at the book every day now. rlhfbook.com/c/11-policy-...

submitted 22 days ago • 2 comments

The real alpha is being the one person working on support vector machines.

submitted 22 days ago • 2 comments

techcrunch.com/2025/02/02/g... #AI #artificialintelligence

submitted 21 days ago • 0 comments

I have to admit that except for the occasional cross-posting, not sure I’ll keep very active on here. There was a big opportunity to migrate the core AI/ML community that was unfortunately squandered with all this hf dataset backlash.

submitted 22 days ago • 33 comments

A lot of researchers think that the "stolen data" claim cope from OpenAI that everyone is now taking at face value is pretty unlikely. Deepseek R1's original paper shows how they're using pure reinforcement learning via GPRO. This is different from previous approaches which human or AI data.

submitted 24 days ago • 0 comments

One big (replicated!) finding on AI doesn’t get enough attention: existing systems like GPT-4 can alter deeply held beliefs using logic & discussion, not manipulation. A short conversation with AI greatly reduce conspiracy theory beliefs (hard to do!) & the effects last months.

submitted 24 days ago • 11 comments

Just curious: will anyone be walking back their conspiracy theories and xenophobic disbelief of DeepSeek’s results, now that Berkeley basically just replicated DeepSeek’s r1?

submitted 24 days ago • 13 comments

𝗖𝗮𝘂𝘀𝗮𝘁𝗶𝗼𝗻, 𝗺𝗮𝗻𝘆 𝘁𝗵𝗶𝗻𝗴𝘀 Always thought causation connected very closely to notion of modularity in general but specifically in neuroscience. Finally found some discussion by Cartwright! (maybe I just missed discussions on this... and ofc there are discussions of causation in complex systems etc)

submitted 24 days ago • 2 comments

It's funny reading responses to my stories where people are like "there is advanced math" Bc I spend a significant amount of time thinking about the huge amount of math I currently don't know, surrounded by ppl who know orders of magnitude more math than I do, so what I know doesn't feel advanced

submitted 24 days ago • 2 comments

Trump announces up to $500 billion in private sector AI infrastructure investment. This is on the scale of Apollo program and Manhattan project. 166 out of 195 countries in the world have a GDP smaller than this investment. www.cbsnews.com/amp/news/tru...

submitted 32 days ago • 0 comments

SambaNova's EvaByte The open-weight tokenizer-free language model. Their 6.5B byte-level LM—-EvaByte matches modern tokenizer-based LMs with 5x less data & 2x faster decoding!

submitted 33 days ago • 2 comments