Dear Connections,

I am thrilled to share my journey in the data field and my passion for AI. With over six years of experience, I have honed my skills in leveraging advanced analytics to improve products and services for customers. Currently, I am the Technical Data Analyst at Sunrise UPC, where I have been instrumental in setting up data analytical tools, creating and maintaining models, and onboarding new data sets. My work has helped the company make data-driven decisions, improving customer experiences.

Before this, I worked as a Data Scientist at IBM, developing machine learning and deep learning models that have helped companies achieve their AI goals. My passion for Speech Recognition and Computer Vision has led me to create Automatic Speech Recognition for the Lithuanian language in my free time.

Apart from my professional experience, I am a Co-founder of the Lithuanian AI Research Group. Through my work, I am committed to advancing the field of AI and empowering the next generation of data professionals.

I am excited to connect with like-minded individuals and explore opportunities to collaborate on cutting-edge projects. Please feel free to reach out if you would like to discuss anything related to data analytics, AI, or anything in between.

Thank you for taking the time to read my profile. Let's connect and stay in touch!
If you catch me outside of work, I'm a passionate runner.🏃‍♂️

I also like meeting new people over a cup of tea ☕. Feel free to reach out at mataciunasdeividas@gmail.com
 
Competencies: Python, SQL, PyTorch, PySpark, Keras, Machine learning, TensorFlow, Business Intelligence, PowerBI, Azure Cognitive Services, Amazon SageMaker

This is crazy! Meta just dropped SAM 3D, you can auto-select any object in a still image and turn it into a high-quality 3D model 🚀

Just click on an object in an image, and it will automatically rebuild the full shape and texture for you.

Two models in the release:
 🛋️ SAM 3D Objects for everyday items and full scenes
 🧑‍🤝‍🧑 SAM 3D Body for human pose + shape

It’s fast, surprisingly accurate, and completely free to try in the playground.

Gemini 3, SAM 3 Subject, SAM 3D... This is a crazy week!!! 🚀

#ai #computervision #robotics #llm

David Mataciunas

Co-Founder @ Stealth Startup | AI researcher

Nano banana Pro!!! 💛💛💛

Another huge release from Google DeepMind Nano banana Pro !!!! Crazy good image model 🚀🤯

First prompt is to create detailed blueprints of Vilnius TV tower!! 🤯🤯🤯


#ai #nanobanana #google #aiimage

Deividas Matačiūnas

I have over 5 years of experience working in the data field. Currently, I work as the Data Scientist at IBM, where I build machine learning and deep learning models that are helping companies achieve their AI goals! I'm deeply passionate about Speech recognition and Computer vision applications. In my free time I creating Automatic Speech Recognition for the Lithuanian language. 🎙
First model using common voice dataset: https://huggingface.co/DeividasM/wav2vec2-large-xlsr-53-lithuanian 🤗

Previously, I worked at INVL Asset Management as a Data Analyst, improving products and services for our customers by using advanced analytics, setting up data analytical tools (i.e. Qlik Sence), creating and maintaining models, and onboarding new data sets.

If you catch me outside of work, I'm a passionate runner.🏃‍♂️

I also like meeting new people over a cup of tea ☕. Feel free to reach out at mataciunasdeividas@gmail.com
 
Competencies: Python, SQL, PyTorch, PySpark, Keras, Machine learning, TensorFlow, Business Intelligence, PowerBI, Azure Cognitive Services, Amazon SageMaker

Google Gemini 3 results are crazy!!! 🤯🤯🤯🤯

I got famous Anthropic thinking hat! 🤖🧢

Visited the Anthropic x AirMail pop-up in London. Great energy, great people, a whole café of AI minds sharing ideas and having fun.

Great Saturday morning! 🚀

#ai #anthropic #london #machinelearning

Nano Banana Pro dabar gali išspręsti net ir valstybinio matematikos brandos egzamino užduotis 🤯

Google naujas vaizdų generavimo modelis ne tik kuria itin realistiškas nuotraukas – jis supranta kontekstą, tekstą skirtingomis kalbomis ir… sprendžia matematinius uždavinius iš nuotraukos.

Norėjau pats tai patikrinti. Pasiėmiau vieną iš matematikos egzamino uždavinių ir paprašiau modelio sukurti vaizdą, kuriame būtų ir užduotis, ir sprendimas.
Ankstesni modeliai dažnai „nesuprasdavo“, kas parašyta nuotraukoje, o čia – priešingai. Nano Banana Pro ne tik perskaitė užduotį, bet ir vizualiai paaiškino sprendimą pačiame sugeneruotame paveiksle.

Ironiška: man patikrinti sprendimą užtruko ilgiau nei modeliui jį sugeneruoti.

AI per kelis mėnesius šoktelėjo į visiškai naują lygį.

#ai #math #gemini #nanobanana

Black Friday is getting a little wild this year, but in a surprisingly good way !!!!

I just came across a really smart campaign from Linkedist where you can stack different “charms” to build your own customized LinkedIn growth package.

If you want better GEO: add a charm.
If you need stronger content: add another.

This year, I chose a GEO charm, or in my case, a gold chain. Generative Engine Optimization is a strategy to make your content and brand more visible in AI-powered search results. I'm glad that I could help the Linkedist team at the beginning of the GEO journey. 

IF you want to improve your game in GEO or in general, get more followers on LinkedIn. Link is in the comments ⬇️

P.S. Nano banana PRO is crazy good for editing :D 

#ai #geo #linkedist #blackfriday

Google just released a new LLM model, GEMINI 3 🚀🤯

The most noticeable aspect is its balanced performance. Gemini 3 Pro excels in reasoning, math, multimodal tasks, coding, and long-context benchmarks, rather than focusing on just one area. The Deep Think mode further enhances reasoning tests, aiding in creating agentic workflows.

Another notable highlight is how Google is rolling out Gemini 3. Instead of a big launch all at once, they’re introducing it step by step through different platforms: the Gemini app, AI Mode in Search, Google AI Studio, Vertex, and a new CLI for developers. Plus, Search now has a more lively, interactive interface powered by the model, hinting at exciting new AI features for users in the future.

Overall, Gemini 3 seems to be Google’s most stable and versatile release yet. It’s not flashy, but it’s reliable, and the benchmark table confirms it.

#ai #llm #gemini3 #deeplearning

This is a fantastic graph showing how China has started leading in open-source in just two years! I believe Europe still has a great chance to catch up, but we need to put more focus on this 🤯🚀

What do you think?

Graph credit: Lukas Ekwueme

#ai #opensource #europe #china

basedspace_ is finishing the year in the best way possible!!! 🤯👽

They’re opening their doors every weekend for the new nites & weekndz series. You show up on Saturday, work alongside 30+ other founders pushing real ideas, and share what you shipped on Sunday. 
No pitches, no fancy agenda, just a solid weekend grind with like-minded founders.

If you keep showing up and making progress, the team notices. That’s how you earn access to the space and eventually go full-time as a basedfounder.

They’re also hosting a short 5-hour hackathon with ElevenLabs on December 11th, with the theme: conversational agents. Global prizes, local prizes, and a very strong crowd.

If you’re in Vilnius and want to end the year by building something that matters, I genuinely recommend joining. Apply and show up. 😎

Ignas Gecevičius jonas bartasius Tomas Dirvonskas Tech Philomaths

#ai #builders #basedspace #elevenlabs

AI progress is real !! Salmon in the river 😂
 
Photo credit: Min Choi

Anthropic released new model Claude Opus 4.5 best at coding and agentic tasks! 

🟠 SWE-bench Verified: Claude Sonnet 4.5 hits 77.2%, up to 82% with high compute.
🟠 OSWorld: Scores 61.4%, a big jump from Opus 4.1’s ~44%.
🟠 Autonomy: Runs 30+ hours in multi-step agent tasks (old models capped at ~7 hours).
🟠 Pricing: Around $5/M input tokens and $25/M output tokens.
🟠 Strengths: Major upgrades in tool-use, agent orchestration, long-context coding, and enterprise reliability.

Last Monday Gemini 3
This Monday Claude Opus 4.5
Who will release next Monday? OpenAI? xAI? or maybe Meta will back to the game 🚀

#ai #claude #anthropic #coding

Ilya Sutskever (co-founder of OpenAI and SSI) was on the Dwarkesh Podcast this week, and it was one of his most interesting conversations in years. A few points stood out:

1. 𝐈𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞 = 𝐭𝐡𝐞 𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐭𝐨 𝐥𝐞𝐚𝐫𝐧, 𝐧𝐨𝐭 𝐭𝐡𝐞 𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐭𝐨 𝐤𝐧𝐨𝐰
Ilya argues that the real milestone is a system that learns from experience during deployment, not one that just absorbs massive pre-training datasets.

2. 𝐑𝐞𝐚𝐥 𝐬𝐚𝐦𝐩𝐥𝐞-𝐞𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐭 𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐧𝐞𝐞𝐝𝐬 𝐚 𝐯𝐚𝐥𝐮𝐞 𝐟𝐮𝐧𝐜𝐭𝐢𝐨𝐧
Dense feedback (TD learning) can guide models even in the absence of explicit rewards. This is closer to how humans learn quickly from limited experience.

3. 𝐓𝐡𝐞 𝐛𝐨𝐭𝐭𝐥𝐞𝐧𝐞𝐜𝐤 𝐢𝐬 𝐚𝐥𝐠𝐨𝐫𝐢𝐭𝐡𝐦𝐢𝐜, 𝐧𝐨𝐭 𝐜𝐨𝐦𝐩𝐮𝐭𝐚𝐭𝐢𝐨𝐧𝐚𝐥
He says we still don’t have algorithms that can learn reliably using similar computing to inference. Those algorithms are essential for continual learning - and likely within reach. A handful of GPUs per researcher is enough to discover them. Giant clusters are only needed once you want to prove to the world that your method scales.

4. 𝐓𝐡𝐞 “𝐞𝐯𝐚𝐥𝐬 → 𝐫𝐞𝐚𝐥 𝐰𝐨𝐫𝐥𝐝” 𝐝𝐢𝐬𝐜𝐨𝐧𝐧𝐞𝐜𝐭
Models look great on benchmarks but lag in impact. Ilya suggests a simple explanation: RL training is optimized for eval scores, not generalization. When eval optimization meets weak generalization, you get impressive numbers but limited real-world robustness.

5. 𝐓𝐡𝐞 𝐚𝐠𝐞 𝐨𝐟 𝐬𝐜𝐚𝐥𝐢𝐧𝐠 𝐢𝐬 𝐞𝐧𝐝𝐢𝐧𝐠
This is the most important statement: more compute won’t give us the next leap. We’re back in a real research era - new ideas, new algorithms, and more diversity of approaches.

6. 𝐇𝐢𝐬 𝐯𝐢𝐞𝐰 𝐨𝐧 𝐬𝐮𝐩𝐞𝐫𝐢𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞
Ilya believes that different technical paths will eventually converge, and that the first true superintelligence must be aligned, democratic, and care for sentient life - “the condition everyone should strive for.”

It’s rare to hear someone this central to AI’s history openly say:
𝘞𝘦’𝘳𝘦 𝘯𝘰𝘵 𝘣𝘭𝘰𝘤𝘬𝘦𝘥 𝘣𝘺 𝘎𝘗𝘜𝘴. 𝘞𝘦’𝘳𝘦 𝘣𝘭𝘰𝘤𝘬𝘦𝘥 𝘣𝘺 𝘪𝘥𝘦𝘢𝘴.

#ai #ssi #agi #ilyasutskever

NeurIPS 2025 just announced this year’s Best Papers 🤯 🚀

NeurIPS is the world’s top conference for AI research, and this year’s selection shows how fast the field is shifting: models becoming more similar than expected, attention getting a small but powerful upgrade, RL going extremely deep, and diffusion models revealing surprising regularization behavior.

Highly recommend reading this 🚀

Winners: 
🏆 Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)
🏆 Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free
🏆 1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities
🏆 Why Diffusion Models Don’t Memorize: The Role of Implicit Dynamical Regularization in Training
🏆 Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
🏆 Optimal Mistake Bounds for Transductive Online Learning
🏆 Superposition Yields Robust Neural Scaling

More information in the blog post above ⬇️

#ai #neurips2025 #machinelearning #deeplearning

The Walt Disney Company reveals their first free-roaming Olaf robotic figure 🤯 This is the most famous snowman in the world, now powered by AI 😂

A free-roaming animatronic that chats with guests, powered by reinforcement learning. He walks, he talks, he vibes.

Coming to Disneyland Paris & Hong Kong Disneyland in 2026.

We officially live in a world where even snowman are getting AI upgrades. 😂

#ai #robotics #disney #llm

Physical Intelligence just showed a robot that learns through experience and can already make espresso, fold laundry, and assemble boxes. The espresso part really stood out to me; barista-level tasks are now becoming repeatable for robots.

It’s not hard to imagine cafés and restaurants letting a system like this handle the daily coffee rush. No drama, no stress, same quality every time. Baristas won’t disappear, but it’s clear that the gap between “robotic arms in labs” and “your morning flat white” is getting very small.

#ai #robotics #llm #deeplearning

Nano Banana Pro infographics are great!!🚀

If you're into AI, check out Pantheon on Netflix. It's wild that Sam Altman recommended it as the best TV show 🤯 - it's really amazing.


#ai #uploadintelligence #deeplearning #tvshow

Anthropic released Interviewer, a tool that runs short real-time interviews to understand how professionals actually think about AI at work.

𝐇𝐨𝐰 𝐢𝐭 𝐰𝐨𝐫𝐤𝐬
The system generates an interview script, talks to users for 10–15 minutes, and sends the transcripts to human analysts. It’s designed to capture honest attitudes, not just usage metrics.

𝐖𝐡𝐚𝐭 𝐩𝐞𝐨𝐩𝐥𝐞 𝐬𝐚𝐢𝐝
⚡ AI saves time and removes repetitive work
✍️ It helps with writing, planning and brainstorming
⚠️ Concerns remain about creativity loss, job identity and long-term displacement
🔬 Scientists and creatives showed the most mixed feelings
🤝 Most want AI as an assistant, not a replacement

𝐖𝐡𝐲 𝐢𝐭 𝐦𝐚𝐭𝐭𝐞𝐫𝐬
Interviewer is one of the few attempts to measure real human sentiment about AI at scale. If you’re building AI products, this is the type of grounded feedback that shows how expectations and fears are shifting.

#ai #anthropic #llm #aiagents

DeepSeek model DeepSeekMath-V2 achieved gold-medal results on the International Mathematical Olympiad (IMO). First open-source model achieves these results!

🥇 5 out of 6 problems solved on IMO 2025 DeepSeekMath_V2
🥇 Gold-level score on CMO 2024
🧠 118/120 on Putnam 2024 — beating the top human score of 90
📊 Strongest performance on ProofBench Basic and highly competitive on Advanced
🔁 A scalable “search + verify + refine” approach that resembles how experts solve problems

#ai #deepseek #math #imo

Deepseek

Why are robots being trained in martial arts? 🤯😂

taplio

David Mataciunas's Linkedin Analytics