dotAI 2026
who it's for
LLM engineers, machine learning engineers, research scientists, MLOps engineers, infrastructure engineers, product engineers, site reliability engineers
topics
mechanistic interpretability and concept vectors; llama.cpp speculative decoding; WebMCP page-exposed tools; retrieval augmented generation (RAG); SFT and RL post-training paths; latent-space hallucination probes; BF16 exponent Huffman compression (Unweight); model distillation and edge compression; agent observability and analytics (BigQuery); production agent design and tooling; benchmarking dynamic AI systems; tabular models for prediction
| title | author | length |
|---|---|---|
| Mechanistic Interpretability | David Louapre | 20m |
Mechanistic interpretability shows that transformers encode identifiable concept vectors rather than single neurons, found by supervised probes or heavy… | ||
| Don't Become Your Agent's B**** | Guillaume Vernade | 21m |
Agents spam noisy pull requests, ask trivial questions, and force constant context switching; model prompts cost a few cents and full runs are about $1, so… | ||
| When Prompting Runs Out: Post-Training an Open Model | Ziv Ilan | 21m |
Maps the cheap levers—newer base checkpoint, context engineering, RAG, tools, constrained decoding, routing—and gives the concrete exit conditions for each… | ||
| AI for Prediction: Why You Need a Statistician Not a Philosopher | Kevin Scaman | 21m |
Strong claim: real-world prediction needs statisticians not philosophers. Simple experiments ask an LLM to predict from 120 train / 30 test rows and show LLM… | ||
| WebMCP: Stop scraping the Web, start calling it | Patrick Brosset | 20m |
Demo shifts from brittle DOM scraping to explicit page-exposed tools: a pizza-order scenario goes from a slow, click-and-screenshot agent to a direct tool call… | ||
| llama.cpp - faster and easier for everyone! | Xuan-Son Nguyen | 11m |
llama.cpp adds speculative decoding to run local inference faster by drafting multiple tokens per decode and then verifying them, using implementations like… | ||
| How to Be Correct at Scale With No Ground Truth | Diane Le Douaron | 09m |
Mirakl Catalog Transformer is an LLM-based pipeline that enforces deterministic correctness for fixed-value lists and required attributes, e.g. mapping dark… | ||
| From Blank Page to Prod: A Year of Building AI Agents at Carrefour | Guillaume Blaquiere | 10m |
Company deploys agents in early 2025 using Gemini and Google’s Agent Development Kit, integrating RAG, observability and an agent-analytics pipeline that… | ||
| Your AI agents have ADHD | Greg Qualls | 11m |
Claims that agent failures map one-to-one to ADHD executive-dysfunction patterns and that clinical coping strategies become deployment checklists. Shows… | ||
| Hallucinations: Harvesting Uncertainty | Aygalic Jara | 10m |
Claims that large models expose a usable internal uncertainty signal readable before generation and token by token, and that token probabilities are too noisy… | ||
| How Do We Compress Giant AI Models to Run at the Edge? | Léo Arsenin | 11m |
Cloudflare describes a lossless compression that targets BF16 exponents by applying Huffman coding to the exponent stream, decompressing directly into on‑chip… | ||
| Radiological AGI: How Close Are We? | Hélène Philippe | 10m |
An agentic pipeline calls a toolbox of mocked tools (simulated perfect detectors and measurers) and is evaluated on the PARROT dataset with the RadMatch… | ||
| Will the future of code be formal or informal? | Stanislas Polu | 22m |
Human attention becomes the bottleneck so agents run faster than humans can review code, prompting a hybrid: keep code but attach small structured… | ||
| How Anthropic Uses Claude | Ian Massingham | 21m |
Anthropic runs Claude across Slack, Drive, email and Salesforce to let agents read context, create code and operate for hours autonomously, reporting 8x… | ||
| The Moving Target: Benchmarking Dynamic AI Systems End-to-End | Gaëtan Brison | 18m |
Doctolib is building a patient-facing agentic stack spanning triage, clinical interview, retrieval, memory, booking and image understanding and evaluates it… | ||
| We Thought the Digital World Would Be Easy to Simulate. Was it? | Maxime Langevin | 21m |
Argues that simulating the digital world is hard in practice but solvable by models that simulate language-form interfaces rather than pixels. Uses VLMs… | ||
| Own your agent | JOFFREY THOMAS, Henry Lagarde | 20m |
Argues that running agents on open weights avoids data exfiltration and vendor lock-in, shows how to build a stack with quantization, fine-tuning, embeddings… | ||
| Beyond the Chatbox: What Changes When AI Runs Where You Are | Christian Keller | 21m |
Argues that device constraints—not model leaderboards—should drive design, showing working examples: live translation in glasses that runs offline, Quest… | ||
| nothing matches | ||