Digital Factory AI Lead

Reshaping industry workflows by developing AI solutions and effective agentic systems.

Lead AI Engineer

moatless

Helping industry experts and researchers to navigate the sheer endless sea of Large Language Model (LLM) research & innovation because everybody deserves to find the best information.

Founder

Stealth

I was part of a 6 month government-funded Entrepreneurship program that enabled me to work on my own projects and ideas within the LLM space.

During this time, I gained experience in building LLM-powered backends with popular in-demand tools. The prototype I ended up building included the following:

- Fetching data from an external API & processing the response (Python)
- NoSQL database (MongoDB, Postgres)
- PDF Processing (PyPDF2, PyMuPDF)
- Document Extraction (Textract/Document Transformer)
- LLM Chain with multiple steps and prompt templates (Python, Guidance/LangChain)
- API endpoint for the processed data & the LLM outputs (FastAPI)
- Cloud Infrastruture (AWS)

Since I operated on a minimal budget, efficiency and cost reduction were a priority for me. Specifically, I focused on optimizing tasks like PDF processing, document understanding, information extraction and summarization.

Senior Machine Learning Engineer

A1 Telekom Austria AG

During my time as a Senior at A1, I was the DS lead for the company's most advanced NLP projects.

My work included

- Reporting results to several levels of management 
- Representing cutting edge Deep Learning & NLP at the company
- Pitching these topics to the C-Level and other influential exexutives
- Giving expert talks & presentations
- Participating in workshops
- Taking part in the recruiting process for full-time Data Science internships
- Guiding selected interns and assigning them roles in my projects

For a technical overview of my work, see below.

Senior Data Scientist

I was in charge of a Social Media Listening project for Facebook, Twitter and Online Newspapers which included the following

- SQL & MongoDB databases
- Web scraper for the newspaper(s)
- Processing Facebook & Twitter API data
- Transfer Learning & Language Model Finetuning on in-domain texts
- Multiclass and Multilabel Text Classifiers based on Embeddings and Transformers (the original PyTorch implementation of BERT)
- Tableau Dashboard

On top of that, I was developing intents & entities for customer text channels, including e-mail, web search, specialized chatbots (e.g., billing), the aforementioned Social Media data and several smaller Text & Sequence Classification projects.

I then moved on to explore speech channels and became the person in charge of Speech-to-Text at the company, which included 

- Processing short IVR audio recordings
- Processing long(er) call center recordings 
- Labeling pipeline with custom Prodigy scripts
- Developing a subtitle matching algorithm for aligning audio with text & filtering out errorenous subtitles
- Developing Hybrid HMM-TDNN Speech Recognition models with Kaldi
- Switching to (back then freshly released) bleeding edge libraries & models for End-to-end Deep Learning, such as NVIDIA NeMo & Facebook FAIR
- Improving performance on Austrian dialects & in-domain recordings
- Adapting the in-house Classification models mentioned earlier for the text output of the speech channels

In addition, I was able to work on PoCs in the areas of Document Extraction, Document Classification, Entity Extraction & Question Answering.

Data Scientist

Universität Wien

Research in Cognitive & Economic Psychology:

- Dual process theory (Type I/II thinking)
- Metacognition theory
- Experimental assessment of cognitive biases
- Leveraging aggregated mouse cursor movement data

Junior Researcher

Coursera

Data Science Education

As the founder of LLM Watch, my goal is to curate the most relevant Large Language Model (LLM) research and make it accessible to a broad audience. 

I distill complex technical topics into summaries that respect your time. 

Before embarking on this journey, I led Deep Learning R&D at one of Europe's largest telecom companies. There, I spearheaded numerous AI projects with a focus on Natural Language Processing (NLP) and Automatic Speech Recognition (ASR). 

With more than 6 years of experience in AI, I was there before the hype. 

I try to offer a healthy balance between shiny new things and what really matters.

Yann LeCun's first big paper after he left Meta. Here's what they did:

For years, training Joint-Embedding Predictive Architectures (JEPAs) has been more art than science. Stop-gradients, teacher-student networks, carefully tuned EMA schedules – a delicate dance of heuristics to prevent representation collapse. But what if there was a principled way?

Self-supervised learning promised to unlock the power of unlabeled data, yet state-of-the-art methods like DINOv2 and I-JEPA rely on complex recipes with dozens of hyperparameters. Each new dataset or architecture requires extensive tuning. Even worse, the training loss often doesn't correlate with downstream performance, making model selection nearly impossible without expensive labeled evaluation.

The core problem: we've lacked a theoretical understanding of what makes good representations. LeJEPA changes this by proving a fundamental result – isotropic Gaussian embeddings uniquely minimize worst-case risk across downstream tasks. 

Tho achieve this, they introduce SIGReg (Sketched Isotropic Gaussian Regularization), a novel objective that enforces this optimal distribution using random projections and characteristic function matching. The beauty lies in its simplicity: linear complexity, bounded gradients, and provably correct by construction.

LeJEPA eliminates stop-gradients and teacher-student networks entirely while achieving competitive performance – 79% on ImageNet-1k with a frozen ViT-H/14. And here's the best part: it works out-of-the-box across 60+ architectures and diverse domains - in-domain pretraining with LeJEPA on small datasets (even just 1,000 samples) outperforms transfer learning from massive frontier models trained on billions of images.

This could fundamentally shift how we think about foundation models. Instead of relying on ever-larger generic models, practitioners in specialized domains can now train competitive models directly on their data with minimal hyperparameter tuning. 

The fact that LeJEPA's training loss actually correlates 99% with downstream performance also solves the critical model selection problem that has plagued SSL research.

What do you think? Is Yann onto something?
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

Pascal Biese

AI Lead at PwC <​/​> Daily AI highlights for 80k+ experts 📲🤗

LeJEPA: Provable and Scalable Self-Supervised Learning

Microsoft just announced the "Era of Agentic Organization". Here's what that means:

As LLMs scale their reasoning capabilities, we need new techniques: sequential thinking is too slow, and parallel thinking wastes compute by running independent traces that can't learn from each other mid-process. The industry needs a paradigm that combines the adaptivity of sequential reasoning with the speed of parallelization.

For this reason, Microsoft Research developed AsyncThink. AsyncThink is a new approach where a single LLM learns to play two roles simultaneously: an organizer that dynamically structures the reasoning process through Fork and Join operations, and workers that execute sub-queries concurrently. 

Their key contribution is that the model learns via reinforcement learning to optimize its own thinking structure - deciding when to split problems, which sub-queries to pursue, and when to merge intermediate results.

The results? 28% lower inference latency compared to parallel thinking while maintaining higher accuracy on mathematical reasoning. Models trained only on simple countdown problems successfully generalize their asynchronous thinking capabilities to completely unseen domains like Sudoku and complex math problems.

The more tasks AI is supposed to solve within an organization, the more important it becomes to structure all of this in a scalable and controllable way. AI agents need to learn to organize their reasoning like skilled project managers, adaptively coordinating multiple thinking processes based on the problem at hand. 
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

Talks about #llms, #genai, #nlproc, #datascience, and #artificialintelligence

Microsoft

The Era of Agentic Organization

Ajay Juneja

NVIDIA did it again - 5-6x faster generation than the current standard.

TiDAR from NVIDIA is a new hybrid architecture that achieves what seemed impossible: 5-6× faster generation than standard autoregressive models while maintaining the same output quality.

The fundamental challenge with current LLMs is computational efficiency. Autoregressive models like GPT generate one token at a time, leaving GPU compute largely underutilized in a memory-bound regime. Diffusion language models can generate multiple tokens in parallel, but this comes at a steep cost - quality degrades significantly when attempting to decode more than one token per step.

TiDAR resolves this tension through a dual-mode architecture. The model "thinks" in diffusion by drafting multiple tokens in parallel, then "talks" in autoregression by sampling final outputs through rejection sampling - all within a single forward pass. 

The key innovation is a carefully designed attention mask that enables both causal and bidirectional attention patterns simultaneously. This allows the model to leverage what the researchers call "free token slots" - additional computation that incurs minimal latency cost in the memory-bound regime.

TiDAR generates 7-8 tokens per forward pass, translating to 4.71x to 5.91x throughput improvements measured in tokens per second. Most importantly, it's the first architecture to close the quality gap with pure autoregressive models while delivering these speedups. The method even outperforms EAGLE-3, currently the leading speculative decoding approach, in both throughput and quality metrics across coding, math, and reasoning tasks.

This could lead to a fundamental shift in LLM inference optimization. Rather than choosing between quality and speed, or relying on separate draft models with limited capacity, TiDAR shows we can achieve both by intelligently combining complementary sampling strategies within a unified architecture.
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

NVIDIA

TiDAR: Think in Diffusion, Talk in Autoregression

Naveen Srivatsav

Can we train billion-parameter models without gradients? Oxford researchers just showed it's possible - at 100x the speed.

Evolution Strategies have always been an elegant alternative to backpropagation. They don't need differentiability, handle noisy objectives gracefully, and parallelize beautifully. But there's been a fundamental problem: they simply don't scale to modern model sizes.

University of Oxford's new EGGROLL algorithm mirrors what made LoRA successful for fine-tuning - instead of perturbing all parameters with full-rank matrices, use low-rank approximations. For each population member, sample matrices A and B to form a rank-r perturbation AB^T rather than a full matrix E. This cuts memory from mn to r(m+n) per layer and reduces forward pass compute proportionally.

What makes this work is subtle but powerful: while each perturbation is low-rank, the final parameter update averages across the entire population, yielding a full-rank update when the population is large enough. The researchers prove this approximation converges to full-rank ES at an O(1/r) rate - faster than typical central limit theorem rates due to the symmetry of the problem.

The results validate the approach across three domains. In pure integer pretraining of recurrent language models, EGGROLL enables stable training with population sizes exceeding 250,000 - orders of magnitude larger than prior work. For LLM reasoning tasks, it matches or exceeds GRPO while maximizing hardware utilization. And critically, it achieves near-inference throughput speeds, making gradient-free optimization genuinely practical at scale.

This opens real possibilities for training systems that can't use backpropagation: end-to-end neurosymbolic architectures, models with discrete components, or deployment scenarios where integer-only computation is essential for efficiency.
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

Evolution Strategies at the Hyperscale

We might be systematically underestimating LLMs.

We've been told that reinforcement learning is essential for frontier reasoning capabilities. Massive compute budgets, curated datasets, reward models - the whole apparatus. 

But new research from Harvard University challenges this expensive paradigm with a surprisingly simple question: can we just sample better from base models?

The paper introduces a training-free algorithm that achieves reasoning performance matching or exceeding GRPO (the standard RL method) across mathematics, coding, and science benchmarks. 

The key insight revolves around the "power distribution" - sampling from p^α instead of p. Unlike low-temperature sampling which greedily averages future token likelihoods, power sampling accounts for entire future completion paths. 

This seemingly subtle difference makes a big difference: it naturally upweights tokens with fewer but higher-likelihood futures, essentially planning ahead rather than getting trapped in dead-ends.

Using Metropolis-Hastings MCMC to approximate this distribution, the researchers show that Qwen2.5-Math-7B can reach 74.8% on MATH500 without any additional training - nearly matching the 78.5% from full RL posttraining. 

On out-of-domain tasks like HumanEval, it outperforms by up to 20 percentage points. Perhaps most remarkably, unlike RL which suffers from diversity collapse, this method maintains strong pass@k performance across multiple samples.

The takeaway is that base models may already contain the reasoning capabilities we're spending enormous resources to extract through RL. This would open up a new axis for inference-time compute scaling that's accessible to anyone with a base model - no training infrastructure, curated datasets, or verifiers required.

We're still learning how to properly query the "intelligence" that already exists in our models.
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

Your Base Model is Smarter Than You Think

Training LLM agents that can actually reason, use tools, and learn from mistakes.

(This has been harder than a lot of people expected.)

A new research framework might change that by finally bringing rigorous reinforcement learning to multi-turn agent training.

The gap between LLMs that generate text and agents that solve problems through environmental interaction is larger than it seems. Traditional RL approaches designed for static, single-turn tasks don't translate well to agents that need to maintain memory, make sequential decisions across multiple turns, and adapt to unpredictable tool outputs.

Agent-R1 is systematically extending the Markov Decision Process framework for agent contexts. The key insight: agents need fundamentally different state representations (capturing full interaction history), stochastic transitions (accounting for environmental feedback), and dense reward signals (process rewards for intermediate steps, not just final outcomes). 

The framework introduces two critical modules - Tool for executing actions and ToolEnv for managing state transitions - and implements action masking to ensure credit assignment focuses on actual agent decisions rather than environmental responses.

Testing on multi-hop question answering, agents trained with Agent-R1 achieved 2.5x better performance than retrieval baselines, with the best configuration (GRPO) reaching 38.77% accuracy. The ablation studies confirm that both loss masking and advantage alignment are essential - removing either causes significant performance degradation.

Agent-R1 rigorously formalizes how RL should work for multi-turn agent training - the modular design means researchers can now experiment with different RL algorithms and task environments without rebuilding infrastructure from scratch.
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

taplio

Pascal Biese's Linkedin Analytics

Pascal Biese

Performance ScorePascal Biese's LinkedIn Performance

What is Pascal talking about?

Who is engaging with Pascal

Pascal Biese's Best Posts (last 30 days)

Ready to grow your LinkedIn presence?

taplio

.css-6ytdur{background-image:linear-gradient(284deg, rgba(39, 43, 72, 1) 0%, rgba(33, 47, 183, 1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Pascal Biese's Linkedin Analytics

Pascal Biese

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Pascal Biese's LinkedIn Performance

What is Pascal talking about?

Who is engaging with Pascal

Pascal Biese's Best Posts .css-1emoo84{font-size:var(--chakra-fontSizes-md);font-weight:400;color:var(--chakra-colors-gray-600);}(last 30 days)

Ready to grow your LinkedIn presence?

Pascal Biese's Linkedin Analytics

Performance ScorePascal Biese's LinkedIn Performance

Pascal Biese's Best Posts (last 30 days)