Researcher in natural language processing, computational linguistics, computational social science
Joined September 2010
- Tweets263
- Following522
- Followers411
- Likes657
Nikola Ljubešić retweeted
Introducing Global PIQA, a new multilingual benchmark for 100+ languages. This benchmark is the outcome of this year’s MRL shared task, in collaboration with 300+ researchers from 65 countries. This dataset evaluates physical commonsense reasoning in culturally relevant contexts.
Nikola Ljubešić retweeted
Oxford researchers just confirmed what we feared:
The internet as we knew it is dying.
AI content went from ~5% in 2020 to 48% by May 2025. Projections say 90%+ by next year.
Why? AI articles cost <$0.01. Human writers cost $10-100.
But the real crisis is model collapse. When AI trains on AI-generated content, quality degrades like photocopying a photocopy. Rare ideas disappear. Everything converges to generic sameness.
It's recursive. Today's AI slop becomes tomorrow's training data, producing worse output, which becomes training data again.
Nikola Ljubešić retweeted
GenAI isn't just a technology; it's an informational pollutant—a pervasive cognitive smog that touches and corrupts every aspect of the Internet. It's not just a productivity tool; it's a kind of digital acid rain, silently eroding the value of all information.
Every image is no longer a glimpse of reality, but a potential vector for synthetic deception. Every article is no longer a unique voice, but a soulless permutation of data, a hollow echo in the digital chamber. This isn't just content creation; it's the flattening of the entire vibrant ecosystem of human expression, transforming a rich tapestry of ideas into a uniform, gray slurry of derivative, algorithmically optimized outputs.
This isn't just innovation; it's the systematic contamination of our data streams, a semantic sludge that clogs the channels of genuine communication and cheapens the value of human thought—leaving us to sift through a digital landfill for a single original idea.
Nikola Ljubešić retweeted
Maybe we should have thought of all these 20 years ago, after China was admitted to WTO and American companies cheerfully laid off blue collar workers & shipped their jobs to China, expecting that they’ll be perfectly happy flipping burgers at McD & greeting shoppers at Walmart?
Nobody bothered to help them out. We heard of Ricardo’s comparative advantage, Schumpeter’s creative destruction, and expected those poorly educated men who lost their good jobs & status to suck it up or get PhDs in computer science and find new jobs. A tax on foreign imports that helped pay for their re-training and supplemented their lost income at least temporarily might have gone a long way. Surely there were other ways to help them. But the laptop class was too busy getting rich and celebrating globalization to notice or care.
Never occurred to anyone that those workers, feeling left behind and disrespected - along with their brethren who still had good paying jobs but saw the writing on the wall and wondered how long till they lose them - would turn to a populist who’d promise to help get their jobs back. And when that populist got elected, never occurred to anyone that he’d actually try to do what he promised.
We will end up a poorer country. And we have ourselves to blame for it, for neglecting our fellow Americans who fell victim to globalization and discovered that the ruling elites couldn’t care less about them.
Nikola Ljubešić retweeted
CLASSLA-Express workshops exploring South Slavic corpora are back!
We hope to see you at the upcoming stops:
📍 4 Apr – Klagenfurt 🇦🇹- registration is already open!
📍 11 Jun – Zagreb 🇭🇷
📍 10 Oct – Graz 🇦🇹
📍 5 Nov – Rijeka 🇭🇷
📍 17 Nov – Bled 🇸🇮
➡️clarin.si/info/k-centre/work…
Nikola Ljubešić retweeted
A silent revolution is happening in Serbia but nobody is talking about it.
It is for justice and against corruption.
It is for peace and against the violence.
It is for the people and against the system.
Please share this and send it to everyone you know 🙏🙏🙏
Nikola Ljubešić retweeted
When we hear someone speak a dialect, we can often tell where they're from. Can LMs do the same?
Our #TACL paper addresses this question and shows how to boost LMs' geolinguistic skills. 🌍
This paper has been in the making for almost three years, so glad it's finally out!
🧵
Nikola Ljubešić retweeted
It seems to me that before "urgently figuring out how to control AI systems much smarter than us" we need to have the beginning of a hint of a design for a system smarter than a house cat.
Such a sense of urgency reveals an extremely distorted view of reality.
No wonder the more based members of the organization seeked to marginalize the superalignment group.
It's as if someone had said in 1925 "we urgently need to figure out how to control aircrafts that can transport hundreds of passengers at near the speed of the sound over the oceans."
It would have been difficult to make long-haul passenger jets safe before the turbojet was invented and before any aircraft had crossed the atlantic non-stop.
Yet, we can now fly halfway around the world on twin-engine jets in complete safety.
It didn't require some sort of magical recipe for safety.
It took decades of careful engineering and iterative refinements.
The process will be similar for intelligent systems.
It will take years for them to get as smart as cats, and more years to get as smart as humans, let alone smarter (don't confuse the superhuman knowledge accumulation and retrieval abilities of current LLMs with actual intelligence).
It will take years for them to be deployed and fine-tuned for efficiency and safety as they are made smarter and smarter.
Nikola Ljubešić retweeted
GPT-4o is our new state-of-the-art frontier model. We’ve been testing a version on the LMSys arena as im-also-a-good-gpt2-chatbot 🙂. Here’s how it’s been doing.
Nikola Ljubešić retweeted
Today at @answerdotai we've got something new for you: FSDP/QDoRA. We've tested it with @AIatMeta Llama3 and the results blow away anything we've seen before.
I believe that this combination is likely to create better task-specific models than anything else at any cost. 🧵
Super excited to finally release the computer-friendly text+speech version of the "Mići Princ" translation of The Little Prince into various Chakavian micro-dialects! hdl.handle.net/11356/1765 Computers like it already, with ASR WER dropping from 35% to 17% and CER from 12% to 4%.
The dataset is also available from huggingface.co/datasets/clas…, and there is already an exemplary Chakavian-understanding whisper-v3-large huggingface.co/classla/whisp…. Joint work with Peter Rupnik and Tea Perinčić! Thank you @PEEKANDPOKE for advocating for this release!
Nikola Ljubešić retweeted
Vabljeni k sodelovanju na konferenci Jezikovne tehnologije in digitalna humanistika #JTDH2024! Konferenca bo gostila govorce s področja jezikovnih tehnologij, digitalnega jezikoslovja in digitalne humanistike.
➡️sdjt.si/wp/jtdh-2024/
🗓️ Rok za oddajo prispevkov: 17. 5.
Nikola Ljubešić retweeted
Excited that our CLASSLA resources for BCS languages were used for such a great cause 👏
Our CLASSLA team is already thrilled to be able to add YugoGPT to the model evaluations on the South Slavic DIALECT-COPA task at VarDial😁Join us or follow along: sites.google.com/view/vardia…
Happy to announce that I'm open-sourcing YugoGPT-base LLM - the best 7 billion parameter LLM for BCS (Serbian, Bosnian, and Croatian) languages! I hope that this contribution of mine will play its part in kicking off the local LLM ecosystem!
You can find the model on HuggingFace: huggingface.co/gordicaleksa/…
If you want to quickly get started with even more powerful models you can use RunaAI's API dev platform here: dev.runaai.com/
AI is the future and the latest paradigm shift in the arc of technological progress.
Countries that fall behind in the AI race are putting their economies, national security, culture & language at risk. It's therefor of utmost importance that every country has a powerful regional LLM tailored to their language(s).
The approach that big tech companies are taking is basically treating most of the world's languages as an afterthought - which is fine for them as it helps them reap some of the profits but not for companies and governments that actually use those languages in production ready systems.
I want to build very powerful regional LLMs and this is just the first step.
A big thank you to all of the project sponsors (listed in the README), and Nikola Ljubešić (@nljubesic), CLARINSI, and CLASSLA for help with BCS data!
Largest open speech+text corpora for Croatian, Serbian and Polish! Rich in metadata (speaker, age, gender, political affiliation and orientation, in power or in opposition etc.).
The ParlaSpeech corpora in Croatian, Serbian, and Polish from the @CLARINERIC #ParlaMint project are now available on CLARIN.SI and @huggingface! They enable exploring parliamentary debates through text AND audio, and development of ASR and TTS for these languages👏
Interested in commonsense reasoning in dialectal texts? The DIALECT-COPA shared task is the perfect fit for you, providing train and dev data for four official South-Slavic languages and two out of three related test dialects sites.google.com/view/vardia… @vardialworkshop @naaclmeeting
The test dialects are the Cerkno dialect of Slovenian (from town of Idrija @visitidrija ), the Chakavian dialect of Croatian (from town of Žminj @ZminjCroatia ), the Torlak dialect of Serbian / Macedonian / Bulgarian (from town of Lebane @opstinalebane ).
Great work by @gordic_aleksa! It was a pleasure to support Aleksa in any way I was able in this high pace endeavour! Looking forward to the continuation of these efforts!
Well, it's official. YugoGPT 7B significantly beats Mistral and LLaMA 2 and is now officially the best open-source LLM in the world for Serbian & other HBS (Croatian, Bosnian, Montenegrin) languages.
Earlier this summer I was frustrated when I saw how poor the situation is as soon as you leave the English NLP space (w/ few exceptions).
Exactly 57 days ago I came up with this idea while working on the Open-NLLB project and training a machine translation system for HBS.
Why not train an LLM for HBS languages? And so yugoGPT was born.
The problem? I had 1 RTX 3090 GPU on my custom built rig. :) (kindly provided to me by @nvidia back in London!)
Fast forward to today I have 16 80 GB A100s from @togethercompute (whom I seriously recommend - not just for their hardware but also for their support & speed & dedication) and 16 more coming in a week (from a different company -> will share more details then).
I had to learn a ton of new stuff as I was basically a single guy (w/ @nljubesic's tremendous help that saved me a ton of time on the data side) building this whole complex system (shout-out to my Discord community was essential for the Serbian LLM eval project).
And I must say: I learned more over the past 6 months than I did during my whole time back at DeepMind. Unless you're very very senior - you just don't get the level of agency I'm enjoying right now.
If I could do this bootstrapped with 16 A100s, give me a cluster of 1000 GPUs and I'll show you how things are getting done :) (hah!) well more on that later, things are cooking in the background.
I'll just say that I'm already working on the 2nd iteration that'll be significantly better.
As for yugoGPT I’ll be red teaming it and I'll need help from the community over the next period to test it. more updates soon :) keep an eye out
Again big thanks to:
* @togethercompute - for generous help with compute
* @tri_dao for connecting me with together! :)
* @wandb - for generous donation of GPT-4 credits for the Serbian LLM eval project (and also in general for their amazing tool that I'm using for logging in this project! life saver!)
* @smarkovic - for spreading the word!!
* @nljubesic - for help around data he's been working on over the past many years
* and to many individuals & companies from the local region (mostly Serbia) that recently supported me!
---
Results obtained via Serbian LLM eval that I've recently released (yesterday - lol): github.com/gordicaleksa/serb…
