Odysseas Zisimopoulos

Meta has recruited 3 researchers from Google DeepMind: Tianhe Yu Cosmo Du Weiyue Wang The trio previously teamed up at Google DeepMind, where they helped build systems that cracked International Mathematical Olympiad (IMO) problems at gold‑medal level. Big things happening for Meta.

5

This paper on the current state of AI Agents is worth reading. Main points: 1. Add memory to agents. 2. Build agents as loops, not pipelines. 3. Go for RL only after the Agent's behavior is reliable. 4. Specify which tool to use when (don't dump 50 tools into a prompt and hope for magic). 5. Multi-agent systems fail unless the roles of planner, executor, and critic are explicit. ------------ Some of these lessons feel obvious, but it's good to know that engineers from Meta, Amazon, and Google agree with each other. Cheers. https://lnkd.in/gncRy-9f

146

6 Comments

We've just integrated Kimi K2 - Moonshot AI's revolutionary open-source model into izzedo chat.  This trillion-parameter agentic specialist redefines what’s possible with AI tool-calling, directly challenging giants like Claude Sonnet 4. Why Kimi K2 Changes the Tool-Calling Game 🧠 1T parameter MoE architecture – 32B activated parameters per query 🛠️ Built for Agentic Workflows: Trained on synthetic data simulating thousands of real-world tool-use scenarios. ⚡️ Claude-Level Execution at 1/4 the Cost: Outperforms Sonnet 4 in tool-calling benchmarks while costing much less. 🔥 Native Tool-Parsing Logic: Handles complex tool invocation autonomously - no manual intervention needed. 🧩 Seamless Multi-Tool Orchestration: Excels at chaining tools for workflows like data extraction → analysis → reporting. Experience Kimi K2 in izzedo chat – available now across all plans. "Kimi K2 represents a leap in agentic intelligence – we're thrilled to make it accessible alongside other leading models." – izzedo engineering team #AITools #AgenticAI #ToolCalling #OpenSourceAI #izzedochat #KimiK2

1

1 Comment

🚀 New Release: EdgeWisePersona Together with Michał Podstawski at TCL Research Europe, we just released EdgeWisePersona — a dataset for evaluating on-device user profiling from natural language interactions in smart homes. 📦 200 users × 50 multi-turn sessions 📊 Benchmarking compact vs. large LLMs 📱 Validated on real smartphones + mobile runtimes 🔗 Dataset: https://lnkd.in/dkZpVtyU 📝 Paper: https://lnkd.in/d5yW89vN 💻 Code: https://lnkd.in/d3pZwqZs Built to support privacy-preserving, long-context personalization at the edge. #AI #LLM #EdgeAI #SmartHome #Dataset #NLP #NeurIPS2025 #TCLResearch

22

New paper on SNNs from my own country #Poland "Three-factor learning in spiking neural networks" by Szymon Mazurek, Jakub Caputa, Jan Argasiński & Maciej Wielgosz published in Cell Press. It has 127 references so it is a great overview of SNNs methods and trends as of 2025. DOI: 10.1016/j.patter.2025.101414

49

2 Comments

[#ICLR2026 𝗢𝗿𝗮𝗹 🎉] 𝗨𝗔𝗟𝗠 (Unified Audio Language Model for Understanding, Generation and Reasoning) has been accepted to #ICLR2026 𝗮𝘀 𝗮𝗻 𝗢𝗿𝗮𝗹. 𝗪𝗲 𝗮𝘀𝗸: Can we build “𝗼𝗻𝗲 𝗮𝘂𝗱𝗶𝗼 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗺𝗼𝗱𝗲𝗹 𝘁𝗼 𝗿𝘂𝗹𝗲 𝘁𝗵𝗲𝗺 𝗮𝗹𝗹”—a model that can understand audio, generate audio, and reason across text+audio—without separate specialized systems? 𝗪𝗲 𝗶𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗲: 1. 𝗨𝗔𝗟𝗠-𝗚𝗲𝗻: a text-to-audio language model that directly predicts audio tokens, with quality comparable to strong diffusion-based approaches. 2. 𝗨𝗔𝗟𝗠: with the right data blending + training/inference recipe, matches specialized SOTA models across audio understanding, text-to-audio generation, and text reasoning. 3. 𝗨𝗔𝗟𝗠-𝗥𝗲𝗮𝘀𝗼𝗻: a multimodal reasoning model that uses both text and audio in intermediate “thinking” steps for complex generation—what we believe is the first demonstration of cross-modal generative reasoning in audio research. 🧠🔊 📄 Paper (OpenReview): https://lnkd.in/gXXWwTrw 📌 arXiv: https://lnkd.in/gh4myyvN 🌐 Project page: https://lnkd.in/gUB7V2KN Work done by an amazing team at NVIDIA, Carnegie Mellon University and University of Maryland: Jinchuan Tian (CMU), Sang-gil Lee, Zhifeng Kong, Sreyan Ghosh (UMD), Arushi Goel, Huck Yang, Wenliang Dai, Zihan Liu, Hanrong Ye, Shinji Watanabe (CMU), Mohammad Shoeybi, Bryan Catanzaro, Rafael Valle, and Wei Ping #ICLR2026 #AudioAI #MultimodalAI #NVIDIA

332

3 Comments

We have a new blog post on using FlexAttention in FlashAttention CuTe DSL! FlexAttention is a popular PyTorch API for implementing many attention variants through custom score modifications ("score mod"), masking and block sparsity ("mask mod"), or both. However, while the original Triton implementation performed well on NVIDIA Ampere GPUs, it significantly lagged behind on Hopper GPUs compared to optimized kernels like FlashAttention-3. On the other hand, although it was possible in principle to add FlexAttention support to the C++ FlashAttention kernels via an EVT-like abstraction, the complexity of CUTLASS C++ and lack of native JIT compilation made this task unnecessarily cumbersome. Porting the FlashAttention kernels over to CuTe DSL therefore presented an ideal opportunity to revisit FlexAttention in FlashAttention. In collaboration with Driss Guessous and Tri Dao, we've now added a simple FlexAttention-style API to FlashAttention CuTe DSL that has minimal performance loss, achieving up to 95% of FA-3 on Hopper, for example. This work is currently available in the official Dao-AILab repo and is also integrated into PyTorch nightly. See the blog for a thorough user's guide on the API: https://lnkd.in/gETPGUzV

394

5 Comments

𝗢𝘂𝗿 𝗽𝗮𝗽𝗲𝗿 𝗼𝗻 𝗙𝗙𝗡 𝗙𝘂𝘀𝗶𝗼𝗻: Rethinking Sequential Computation in LLMs - 𝘄𝗮𝘀 𝗮𝗰𝗰𝗲𝗽𝘁𝗲𝗱 𝗮𝘀 𝗮 𝙎𝙥𝙤𝙩𝙡𝙞𝙜𝙝𝙩 to #NeurIPS2025 !🎉🎉🎉 Developed by the Deci team at NVIDIA, it fuses sequences of FFN layers, trimming depth to make models faster. It powered the base model for NVIDIA’s Llama-Nemotron-Ultra-253B. Paired with Puzzle, we saw up to 1.71× lower latency and 35× lower per-token cost vs the original model. At its April 2025 release, Nemotron-Ultra-253B ranked No. 1 among open-weight models on Artificial Analysis’s Intelligence Index. Join our session at NeurIPS in San Diego to talk about FFN Fusion and practical LLM architecture search. Read it here: https://lnkd.in/deEPAvjm

113

3 Comments

A new Trend with voice Cloning The old way for voice cloning in TTS relied heavily on diffusion based generation, where a model learned to reconstruct speech by reversing a noise process over multiple timesteps. In such systems the model iteratively denoises a noisy representation. But modern architectures condition TTS models directly on speaker embeddings or acoustic prompt tokens, extracted from a short reference audio clip. These embeddings encode speaker characteristics (timbre, pitch distribution, articulation style) and are provided as conditioning inputs along with text to an autoregressive transformer that generates discrete speech tokens. The generation process becomes a straightforward AR prediction of discrete acoustic tokens, which are later decoded into waveform by a neural decoder. As a result, the computational complexity drops from O(NT) to O(N) (where N is the number of denoising steps and T the audio length) But aren't NAR processes faster than AR especially when modeling dense continuous fields like audio waveforms? Diffusion voice cloning generates speech by repeatedly refining a noisy audio signal over many denoising steps. Newer TTS systems condition an AR model on text and a speaker embedding (or acoustic prompt) and directly predict discrete speech tokens. Because the model produces the acoustic representation directly instead of iteratively refining noise, token-based generation typically achieves much lower inference latency than diffusion-based approaches.

18

2 Comments

The Triton team at Meta has a working implementation of warp specialization in the Triton compiler, presented at the Triton and PyTorch conferences in the past year. This post outlines the current design of warp specialization (autoWS) in Triton and discusses the upcoming roadmap. The team is beginning to build this work with the open source Triton community and industry partners and invites feedback on the design and roadmap. 🔗 Read our latest blog: https://lnkd.in/ggGeEkTV ✍️ Partha Kanuparthy, Nicholas Riasanovsky, Neil Dhar, Jie Liu, Hongtao Yu, Shane Nay, Manman Ren #PyTorch #Triton #AIInfrastructure #OpenSource

120

1 Comment

🚀 Moving AI research from theory to real-world impact At ICLR this year, Shaked Meirom, Yonatan Belinkov, and Erez Safahi presented AI21 Labs' paper 'Jamba: Hybrid Transformer-Mamba Language Models'. It was exciting to connect with so many researchers already using Jamba in their own work and to share how AI21 turned the once-theoretical concept of hybrid architecture into market leading, production-ready models. The importance of bridging theoretical AI research with real-world production applications was a major theme at this year's conference, and we were proud to showcase how AI21 is leading the way. The AI21 Jamba family of open models are available today on Hugging Face. Building something with AI21’s Jamba? Let us know in the comments! AlphaSignal #ICLR2025 #AIResearch #LLMs

129

1 Comment

𝗢𝘂𝗿 𝗽𝗮𝗽𝗲𝗿 𝗼𝗻 𝗙𝗙𝗡 𝗙𝘂𝘀𝗶𝗼𝗻: Rethinking Sequential Computation in LLMs - 𝘄𝗮𝘀 𝗮𝗰𝗰𝗲𝗽𝘁𝗲𝗱 𝗮𝘀 𝗮 𝙎𝙥𝙤𝙩𝙡𝙞𝙜𝙝𝙩 to #NeurIPS2025 !🎉 Developed by the Deci team at NVIDIA, it fuses sequences of FFN layers, trimming depth to make models faster. It powered the base model for NVIDIA’s Llama-Nemotron-Ultra-253B. Paired with Puzzle, we saw up to 1.71× lower latency and 35× lower per-token cost vs the original model. At its April 2025 release, Nemotron-Ultra-253B ranked No. 1 among open-weight models on Artificial Analysis’s Intelligence Index. Join our session at NeurIPS in San Diego to talk about FFN Fusion and practical LLM architecture search. Read it here: https://lnkd.in/dF5mFG5Z

36

1 Comment

Spiking neurons meet State Space Models: 🚨 Our new preprint is out! With my brilliant co-author Lyubov Dudchenko and supervisor Emre Neftci, we introduce a new class of spiking neurons that leverage the structure and training advantages of SSMs (State Space Models). Since the emergence of the AdLIF (adaptive leaky integrate and fire) neuron for speech recognition, SNN progress has slowed - especially wihtout the use of costly synaptic delays. Our work changes that. 🔀 Enter SSMs: Models like S4, have shown impressive long-sequence modeling through clever use of control theory and training-friendly structure. We bridge SSMs and SNNs, proposing two new neuron models: 1. SiLIF (SSM-inspired LIF): Enhances the AdLIF with continuous space logarithmic reparametrization and trainable timesteps. 2. C-SiLIF (complex SiLIF): Introduces complex-value dynamics, builds on SiLIF, and integrates S4-style initialization. 🚀 Key results: * Outperforms previous SNNs on speech recognition – no synaptic delays required * Maintains ultra-low compute cost * Shows the benefits of timestep training + reparametrization for stable, efficient SNN learning * Better efficiency–accuracy trade-off than event-based SSMs 💡 We believe this opens new pathways for efficient, high-performance SNNs across domains – with no new inference implementation required. 📄 Paper: https://lnkd.in/eUh6-5Ws 💾 Code: https://lnkd.in/eTyY8qEe #DeepLearning #NeuralNetworks #SpeechRecognition #StateSpaceModels #SpikingNeurons #Reparametrization #BackpropagationThroughTime #SSMs #SNNs

167

5 Comments

The AI inference race just got more interesting. A new benchmark shows AMD MI355X outperforming NVIDIA B200 on FP8 DeepSeek disaggregated inference. The headline will grab attention. But the bigger signal is underneath it. AI infrastructure competition is shifting away from peak GPU specs and toward stack-level efficiency. FP8 optimization. Disaggregated inference architectures. Tokens per dollar at scale. The real question is no longer which GPU is fastest. It’s which platform runs production workloads most efficiently.

13

PINA is now officially part of the PyTorch ecosystem. It’s a clean, modular SciML framework built directly on PyTorch + Lightning: • Unified workflow: problem → model → solver → training • Handles tensors, domains, graphs (PyTorch Geometric ready) • Built-in differential operators, physics losses, soft constraints • Scales out-of-the-box with Lightning (multi-GPU, distributed) Perfect for: • Biomechanics & tissue modeling • Blood flow / CFD surrogates • Drug diffusion PDEs • Deformable organs or prosthetics • Any physics-informed patient-specific simulation Install: pip install pina-sciml GitHub: https://lnkd.in/g89ZXiiG If you ever need physics-informed models in medicine, this is worth a look. #PyTorch #MONAI #HIPAA #AIinHealthcare #HealthAI

1

Training large-scale Mixture-of-Experts (MoE) models efficiently is hard. NVIDIA NeMo Automodel makes it easy with accelerated performance, production-ready recipes and reproducible benchmarks for popular MoE architectures. Learn how this open-source library combines native PyTorch parallelisms with optimizations like NVIDIA Transformer Engine and DeepEP. Get the full details and quick-start guide: https://lnkd.in/gUNxJb2H #NVIDIA #NeMo #PyTorch #DeepLearning #LLM #MixtureOfExperts #AIResearch #OpenSourceAI

123

2 Comments

Loved the “Pydantic for LLM Workflows” course from DeepLearning.ai! This course is an incredibly practical deep dive into how structured data modeling and type validation can make LLM-powered systems far more reliable and predictable. I remember when I prototyped an AI assistant chatbot for our analyst team — I had to spend a lot of time on prompt engineering to ensure the LLM output was structured correctly for downstream reuse. Still, around 10–20% of the time, the model output contained extra natural language wrapping — things like “Here is the result in JSON format:” or even extra quotation marks — which caused the pipeline to break. With Pydantic, we can safeguard LLM outputs, automatically validate the response schema, and even re-iterate until we get the expected structured output — which I found truly fascinating! The course explored how Pydantic fits into the modern LLM application stack, acting as the backbone for structured input/output control in tools like LangChain and OpenAI function calling. Key technical takeaways: • Use BaseModel to define schemas for LLM inputs and outputs — ensuring consistency across runs and prompts. • Leverage model_validate() and model_dump() to parse JSON responses into typed Python objects — no more brittle string parsing. • Build multi-step pipelines where each stage (prompt → LLM response → validation → downstream task) uses Pydantic models for data integrity. • Apply custom validators to catch malformed or incomplete responses early. • Integrate Pydantic with LangChain/OpenAI to serialize and deserialize structured data seamlessly. In essence, this course bridges LLM creativity with software engineering rigor — a skillset that’s becoming crucial as we build AI systems interacting with real business logic and sensitive data. Excited to apply these concepts in my work around business fraud detection and identity verification pipelines, where reliability and consistency are just as important as intelligence. Thanks, Ryan Keenan! #LLM #Pydantic

21