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.

Pascal Biese

Data scientists spend up to 80% of their time manually crafting features. What if an agent could do it for you?

FAMOSE is a new framework that uses the ReAct paradigm to automatically discover, generate, and select features for tabular data.

Think of it like giving an ML engineer a reasoning loop: the agent observes the data, thinks about what features might help, acts by generating them, then evaluates whether they actually improved performance.

The key insight: instead of brute-forcing through an exponentially large feature space, FAMOSE explores strategically. It generates candidate features, tests them against model performance, and iteratively refines its approach based on what works.

This matters because feature engineering has always required deep domain expertise. You needed to know that "time since last purchase" matters for churn prediction, or that "ratio of income to debt" predicts loan defaults.
FAMOSE shifts that burden from human intuition to autonomous exploration.

For practitioners working with tabular data, this could dramatically reduce the manual iteration cycle. 

The question worth asking: how well does automated feature discovery generalize across domains where human intuition currently dominates?
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

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

FAMOSE: A ReAct Approach to Automated Feature Discovery

19.5 FPS video generation on a single H100.

That's real-time. From a 14B parameter model.

Conventional wisdom said you needed KV-cache, sparse attention, or quantization to hit these speeds. You also needed anti-drifting heuristics like self-forcing or keyframe sampling to prevent long videos from degrading into visual noise. 

Helios throws out both assumptions. The team built a 14B autoregressive diffusion model that generates minute-scale videos without the standard acceleration toolkit. Instead of bolting on complexity, they went the opposite direction: heavy compression of historical context, fewer sampling steps, and infrastructure-level optimizations that cut memory so aggressively you can fit four 14B models in 80 GB of VRAM.

For the drifting problem, they characterized exactly how videos fall apart over time and built training strategies that simulate drift explicitly. The model learns to recover from its own mistakes rather than relying on runtime heuristics to mask them.

The computational cost ends up comparable to 1.3B models - roughly 10x smaller architectures.

On both short and long video benchmarks, Helios outperforms prior methods while running at speeds that actually matter for production. Code, base model, and distilled model are all coming.

Real-time, minute-scale video generation without the usual engineering overhead. The tradeoffs just shifted.
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

 Helios: Real Real-Time Long Video Generation Model

Simple factual questions don't need chain-of-thought reasoning.

That's been the assumption. Why would a model need to "think step by step" to recall a single fact?

New research indicates that this assumption is wrong.

Enabling reasoning on simple, single-hop factual questions substantially expands what a model can recall from its own parameters. Facts that are effectively unreachable through direct answering become accessible when the model reasons first.

The researchers identified two mechanisms driving this:

First, a computational buffer effect. The model uses generated reasoning tokens to perform latent computation - regardless of what those tokens actually say. The extra token generation acts as processing room, giving the model more compute to locate the right answer.

Second, factual priming. When the model generates topically related facts during reasoning, those facts act as a semantic bridge to the correct answer. It's self-retrieval: the model primes itself by producing nearby knowledge, then follows that trail to what it actually needs.

The big problem? When the model hallucinates intermediate facts during reasoning, the final answer becomes more likely to hallucinate too. The same priming mechanism that helps retrieval can poison it.

Filtering reasoning trajectories for hallucination-free intermediate statements directly improved accuracy. If you're using reasoning-enabled models for knowledge-intensive tasks, the quality of the chain-of-thought matters - not just its presence.
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

How Reasoning Unlocks Parametric Knowledge in LLMs

Paul Iusztin

🤖 Senior Machine Learning Engineer • MLOps • Contractor ~ Helping startups engineer production-ready ML/AI systems.

I'm the author of the bestseller LLM Engineer's Handbook, lead instructor of the Agentic AI Engineering course, founding AI Engineer of a San Francisco start-up, and obsessed with making knowledge accessible through AI.

With over 10 years of experience in AI and software and 20 apps shipped, I teach AI Engineering as I wanted to at the beginning of my career. End-to-end. From idea to production. From data collection to deploying, monitoring and evaluation. With a focus on AI principles, software patterns and infrastructure systems that will thrive in a future dominated by AI coding tools.

My ultimate goal is to help other engineers escape PoC purgatory and 10x their AI Engineering skills.

.

If you want to become an AI Engineer pro, join Decoding AI for exclusive content on designing, building, and shipping AI software that works: 

🗞️ https://www.decodingai.com

See more:

🌐 Homepage: https://www.pauliusztin.ai
📚 LLM Engineer's Handbook: https://www.amazon.com/LLM-Engineers-Handbook-engineering-production/dp/1836200072/
🤖 Agentic AI Engineering Course: https://academy.towardsai.net/courses/agent-engineering?ref=b3ab31

.

✉️ For collabs, contact me at: p.b.iusztin@gmail.com

Senior Machine Learning Engineer

🟡 Scalable ML batch infrastructure for continuous training & inference

The client, a generative AI leader, struggled with manual training and calling their ML models. As a tedious and unscalable process, they needed to automate their ML processes completely and integrate them with their current web platform.

➞ 𝐃𝐞𝐬𝐢𝐠𝐧𝐞𝐝 and 𝐢𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐞𝐝 a modular 𝐏𝐲𝐭𝐡𝐨𝐧 𝐩𝐚𝐜𝐤𝐚𝐠𝐞 that composes the ML model required steps into 3 main batch pipelines that can easily be automated: feature, training, and inference, 𝐞𝐥𝐢𝐦𝐢𝐧𝐚𝐭𝐢𝐧𝐠 a 𝐭𝐞𝐝𝐢𝐨𝐮𝐬 𝟐-𝐡𝐨𝐮𝐫 𝐦𝐚𝐧𝐮𝐚𝐥 𝐩𝐫𝐨𝐜𝐞𝐬𝐬/𝐯𝐢𝐝𝐞𝐨.
➞ 𝐃𝐞𝐩𝐥𝐨𝐲𝐞𝐝 a scalable & cost-effective 𝐛𝐚𝐭𝐜𝐡 𝐚𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞 and 𝐂𝐈/𝐂𝐃 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 on top of AWS ECS & SQS that scales in and out based on the number of jobs, 𝐬𝐚𝐯𝐢𝐧𝐠 𝐭𝐡𝐨𝐮𝐬𝐚𝐧𝐝𝐬 𝐨𝐟 𝐝𝐨𝐥𝐥𝐚𝐫𝐬/𝐦𝐨𝐧𝐭𝐡.
➞ 𝐒𝐚𝐯𝐞𝐝 𝐦𝐨𝐧𝐭𝐡𝐬 𝐨𝐟 𝐟𝐮𝐭𝐮𝐫𝐞 𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐦𝐞𝐧𝐭 by using good software practices to make the Python module & batch cloud infrastructure flexible and composable.
➞ 𝐁𝐮𝐢𝐥𝐭 a 𝐯𝐨𝐢𝐜𝐞 𝐬𝐲𝐧𝐭𝐡𝐞𝐬𝐢𝐬 𝐏𝐨𝐂 in ~1 week required to 𝐜𝐥𝐨𝐬𝐞 the 𝐝𝐞𝐚𝐥 with one of the 𝐜𝐮𝐫𝐫𝐞𝐧𝐭 𝐜𝐥𝐢𝐞𝐧𝐭𝐬.

⚫️ Big data ingestion & archiving pipeline

The client wanted to expand its platform features by securely and safely ingesting, preprocessing, and storing TB of data into their platform.

➞ 𝐂𝐨𝐧𝐭𝐫𝐢𝐛𝐮𝐭𝐞𝐝 to the 𝐝𝐞𝐬𝐢𝐠𝐧 and 𝐢𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧 of a 𝐝𝐚𝐭𝐚 𝐢𝐧𝐠𝐞𝐬𝐭𝐢𝐨𝐧 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 that supports 𝐢𝐧𝐠𝐞𝐬𝐭𝐢𝐧𝐠 >𝟓 𝐓𝐁 of data in 𝐨𝐧𝐞 𝐠𝐨 in a reliable, fast, and secure way.
➞ 𝐈𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐞𝐝 a scalable 𝐝𝐚𝐭𝐚 𝐩𝐫𝐞𝐩𝐫𝐨𝐜𝐞𝐬𝐬𝐢𝐧𝐠 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 that supports multiple media types (images, videos, audio, 3D objects) and creates small preview files that are cheap to transfer, store and visualize. The 𝐨𝐫𝐢𝐠𝐢𝐧𝐚𝐥 𝐟𝐢𝐥𝐞𝐬 are 𝐚𝐫𝐜𝐡𝐢𝐯𝐞𝐝 on an affordable storage solution 𝐫𝐞𝐝𝐮𝐜𝐢𝐧𝐠 𝐜𝐨𝐬𝐭𝐬 𝐛𝐲 𝐮𝐩 𝐭𝐨 𝟖𝟏%.

I built a system that turns a single idea into a publish-ready technical article.

Here's a mini masterclass on how (so you can copy):

The system consists of 2 parts:
 
• Nova (research agent): exploratory, steerable
• Brown (writing workflow): deterministic, auditable

𝟭) 𝗡𝗼𝘃𝗮: 𝗦𝘁𝗲𝗲𝗿𝗮𝗯𝗹𝗲 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗔𝗴𝗲𝗻𝘁 (𝗲𝘅𝗽𝗹𝗼𝗿𝗮𝘁𝗶𝗼𝗻)

Goal: ingest context from messy sources, fill gaps, and produce a clean research artifact.

Inputs:
 
• Structured article brief (topic, audience, structure, constraints, candidate sources)

How it runs:
 
• MCP-first design so tools are portable across harness clients
• A prompt exposed as procedural memory to plan and glue all stages
• Tooling focused on ingesting context, not “thinking in tokens”

Core stages:
 
• Pull context from URLs, local files, GitHub repos, YouTube transcripts
• Run multi-round research loops (Perplexity for cited web search)
• Filter for trust and relevance
• Deep scrape the top sources
• Consolidate into a single handoff artifact: research.md

Output:
 
• A curated research document with organized notes and citations, ready for drafting

𝟮) 𝗕𝗿𝗼𝘄𝗻: 𝗥𝗲𝗹𝗶𝗮𝗯𝗹𝗲 𝗪𝗿𝗶𝘁𝗶𝗻𝗴 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄 (𝗲𝘅𝗲𝗰𝘂𝘁𝗶𝗼𝗻)

Goal: turn research into a consistent article, with quality gates and checkpoints.

Why a workflow, not an agent?

Because writing quality demands:
 
• Repeatability
• Checkpoints
• Clear stage boundaries
• Traceability when something is off

How it runs:
 
• LangGraph for stateful orchestration, resumability, and auditability
• Leveraging tools as deterministic workers instead of exposing them to agents

Core stages:
 
• Load context: guideline + research.md + writing profiles
• Generate required assets in parallel (diagrams, images, tables when needed)
• Draft the article
• Run review and rewrite loops (reflection scores, style judge, revision triggers)
• Support human edits without breaking the system’s rules

Outputs:
 
• Final article
• Generated assets + metadata
• Reflection scores and traces (so you debug quality instead of guessing)

Here’s the gist…

If you promise “production-grade writing”, you need exploration for research and durability for execution.

This design choice is the difference between a cool demo and a system you can ship.

Want to learn how to build it out yourself?

We break this down step-by-step in the Agentic AI Engineering course (built with Towards AI, Inc.):

Enroll here: https://lnkd.in/dtxZuZpb

P.S. We’ve opened up another 100 seats for $499. The last batch sold out in under 14 days → get in early to lock in our best offer.

Senior AI Engineer • Founder @ Decoding AI • Author @ LLM Engineer’s Handbook ~ I ship AI products and teach you about the process.

taplio

The Battle

Pascal Biese

Paul Iusztin

Followers

Posts

Engagements

Likes

Performance ScorePascal Biese's LinkedIn Performance

Performance ScorePaul Iusztin's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Pascal Biese

Paul Iusztin

Followers

Posts

Engagements

Likes

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

Performance ScorePaul Iusztin's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScorePascal Biese's LinkedIn Performance