Mistral

The Desktop Frontier — Ahmad Osman, Osmantic

Jul 21, 2026 · 18:02

Ahmad Osman, founder of Osmantic, argues that within roughly 18 months (by late 2027) a single RTX 5090 will run intelligence equivalent to GLM 5.2, driven by the Densing Law of increasing impact per parameter. He shows this trend through concrete examples: a 27B-parameter Qwen 3.5 now beats the 405B LLaMA 3, and the same eight RTX 3090s that once struggled with LLaMA 2 can now run 15 parallel Qwen 3.5 agents. Osman presents the Densing Law—every 3.5 months, 50% fewer parameters achieve the same capability—as a systematic pattern, not coincidence. He advocates for sovereign AI: owning your own hardware (like a DGX Station or RTX 5090) gives you control, avoids cloud limitations, and sees hardware appreciate in utility as models become more efficient. He asks why fund cloud data centers when local hardware can run frontier intelligence and grow more valuable over time.

Lessons from Trillion Token Deployments at Fortune 500s — Alessandro Cappelli, Adaptive ML

May 12, 2026 · 18:35

Alessandro Cappelli, co-founder of Adaptive ML, argues that 95% of GenAI pilots fail to reach production because they rely on proprietary models or instruction fine-tuning, which lack systematic feedback integration. Reinforcement learning (RL) is the only post-training technique that mathematically incorporates defects, business metrics, and production signals to continuously improve models. RL enables smaller, cheaper, faster models that enterprises can own—critical for scaling use cases like AT&T’s transcript summarization or Manulife’s agents. For agents, RL naturally fits because it was designed for environments; synthetic data is generated as a byproduct of environment training, not a prerequisite. Reward signals come from business KPIs (e.g., containment rate) or LLM judges defined by human rubrics in hours, not weeks. Adaptive ML’s Adaptive Engine abstracts away RL complexity (orchestrating four models for PPO) and provides pre-built recipes to industrialize model deployment.

POC to PROD: Hard Lessons from 200+ Enterprise GenAI Deployments - Randall Hunt, Caylent

Jul 23, 2025 · 19:16

Randall Hunt from Caylent shares hard lessons from over 200 enterprise GenAI deployments, arguing that evals, embeddings, and prompt engineering matter far more than fine-tuning. He emphasizes that speed and UX are critical; a slow inference kills adoption, while techniques like generative UI and caching can mitigate latency. Hunt details real-world examples: using audio amplitude spectrographs for sports highlight reels, pooling multimodal embeddings for nature footage search, and noting that nurses prefer chat over voice bots in noisy hospitals. He reports zero regressions moving from Claude 3.7 to 4, and advises optimizing context and economics, such as leveraging Amazon Bedrock batch for 50% cost reduction. The talk underscores that knowing your end customer and minimizing irrelevant context are key to production success.

Recsys Keynote: Improving Recommendation Systems & Search in the Age of LLMs - Eugene Yan, Amazon

Jul 16, 2025 · 20:54

Eugene Yan's keynote presents three innovations for recommendation systems: Semantic IDs, LLM-augmented data, and unified models. Kuaishou’s trainable multimodal Semantic IDs increased cold-start coverage by 3.6% and velocity by 3.5% by clustering content embeddings. Indeed used GPT-4 fine-tuning and distillation to filter bad job recommendations, reducing bad recs by 20% while boosting application rate 4% and cutting unsubscribes 5%. Spotify’s LLM-generated exploratory search queries drove a 9% increase in exploratory queries for new categories like podcasts. Netflix’s Unicorn unified ranker matched or exceeded specialized models across search and recommendations, while Etsy’s unified embeddings with a quality vector achieved a 2.6% sitewide conversion lift and 5% more search purchases.

GPU-less, Trust-less, Limit-less: Reimagining the Confidential AI Cloud - Mike Bursell

Jun 3, 2025 · 43:41

Mike Bursell of Super Protocol argues that Confidential AI, built on hardware Trusted Execution Environments (TEEs) like Intel TDX, AMD SEV-SNP, and NVIDIA GPU TEEs, solves the trust problem in AI by enabling secure processing of sensitive data and proprietary models without exposure. Super Protocol, a decentralized confidential AI cloud and marketplace, allows users to deploy models in TEEs, verify execution via cryptographic attestation, and collaborate across organizations without blind trust. Demos show deploying DeepSeek on H100 GPUs, running n8n healthcare workflows, distributing vLLM inference across four GPU nodes, and provably training a medical model on datasets from Alice's lab and Bob's clinic. Case studies include Realize achieving 75% accuracy and 3-5% sales increase for Mars, and BEL reducing FDA audit time from weeks to 1-2 hours. The protocol replaces trust with on-chain proofs, enabling GPU-less, trustless, limitless AI.

Personal, Local, Private AI Agents: Soumith Chintala

Apr 6, 2025 · 20:32

Soumith Chintala, co-founder of PyTorch, argues that personal AI agents should run locally and privately to maintain trust and control over intimate data. He warns that cloud-based agents, lacking complete context (like access to all messaging or financial accounts), become unreliable and potentially dangerous—catastrophic actions like buying a Tesla instead of Tide Pods are possible. Key technical challenges include slow local inference, immature open-source computer-use models, and poor catastrophic action classification. He is bullish on open models surpassing closed ones through coordinated improvement, citing Linux, Llama, and DeepSeek. Chintala also plugs open reasoning data from gr.ink and PyTorch’s work on enabling local agents, urging AI engineers to tackle these gaps.

Lessons from building GenAI based applications — Juan Peredo

Feb 22, 2025 · 33:13

Juan Peredo details the hidden complexities of building GenAI applications, from model hosting and cost control to output validation and observability. He compares local (Ollama) vs cloud hosting (Modal, SkyPilot) and warns that an agent processing 3,000 calls/day with OpenAI O1 costs nearly $300,000/month, while LLaMA 3.3 70B drops that to $50,000/month. He explains techniques to mitigate hallucinations—prompt engineering, guardrails, RAG, and fine-tuning—each with trade-offs like added latency or cost. Peredo advocates externalizing prompts via LangChain Hub for easy iteration and future-proofing, and illustrates agent design with parallel calls to reduce latency. Finally, he stresses observability using tools like LangSmith to debug probabilistic failures, such as an LLM failing on case sensitivity.

LLM Quality Optimization Bootcamp: Thierry Moreau and Pedro Torruella

Feb 8, 2025 · 53:05

Thierry Moreau of OctoAI demonstrates how to fine-tune Llama 3 8B on a PII redaction task using OpenPipe and OctoAI, achieving 47% better accuracy and a 200x cost reduction (from $30 to $0.15 per million tokens) compared to GPT-4 Turbo. He explains that fine-tuning should follow prompt engineering and RAG, and works best for specialized tasks like function calling. The talk walks through building a fine-tuning dataset from the PI Masking 200k dataset, using OpenPipe to train a LoRA for $40, deploying it on OctoAI, and evaluating it to show the fine-tuned model scores 0.97 accuracy versus GPT-4’s 0.68. Moreau emphasizes that this continuous deployment cycle requires monitoring data drift and retraining, but tools like OpenPipe and OctoAI make it accessible even for teams without deep ML expertise.

[Full Workshop] Llama 3 at 1,000 tok/s on the SambaNova AI Platform

Feb 7, 2025 · 1:00:58

Michelle Matern and Petro Milan of SambaNova present their full-stack AI platform, built on the SN40L RDU chip with a three-tiered memory architecture capable of storing up to 5 trillion parameters. They demonstrate Samba1 Composition of Experts (CoE), a trillion-parameter model combining 92 expert models behind a single endpoint, and show Llama-3-8B achieving 1,000 tokens per second with a time-to-first-token of 0.09 seconds and total inference time of 0.65 seconds—far exceeding GPU-based providers. The workshop includes a hands-on basic inference call using LangChain and SambaStudio API, and a RAG-based Q&A system for enterprise search that integrates Unstructured for document loading, E5-large-v2 embeddings, ChromaDB vector store, and the high-speed Llama-3 endpoint. Attendees learn to configure prompts with special Llama-3 tags, set chunk size and overlap, and optionally run embeddings on SambaNova's RDU hardware for faster processing.

Mastering LLM Inference Optimization From Theory to Cost Effective Deployment: Mark Moyou

Jan 1, 2025 · 33:39

NVIDIA solutions architect Mark Moyou explains that LLM inference differs fundamentally from standard deep learning deployment, requiring careful management of KV Cache, attention mechanisms, and GPU memory to control cost. He details how tokens are processed: prefill computes attention across the entire prompt, then generation produces one token at a time, with KV Cache storing key-value pairs to avoid recomputation. Llama's 32 attention heads and FP8 quantization (halving memory with near-identical accuracy) are cited as key optimizations. Moyou emphasizes measuring time to first token, inter-token latency, and input/output sequence length distributions to size inference engines. He presents NVIDIA's TRT-LLM (model compilation for LLMs) and Triton inference server as tools to maximize throughput, and discusses how query patterns like long-input-short-output or short-input-long-output impact GPU utilization and deployment cost.

AI Frontiers in Trust and Safety Combatting Multifaceted Harm on Tinder at Scale: Vibhor Kumar

Dec 2, 2024 · 14:36

Vibhor Kumar, senior AI engineer at Tinder, explains how the company uses open-source LLMs and LoRAX to detect a long tail of trust and safety violations at global scale. Facing challenges like content pollution and automated fraud from generative AI, Tinder leverages pre-trained models such as LLaMA and Mistral, fine-tuning them with LoRA and QLoRA on hybrid datasets generated by GPT-4 and manually verified. They serve dozens of fine-tuned adapters on a single GPU using LoRAX, achieving real-time inference (tens of QPS, ~100ms latency) for categories including hate speech, pig butchering scams, and underage users. The approach yields near 100% recall on simpler tasks and significant improvements over baselines, with better generalization that resists adversarial evasion. Future directions include visual language models for explicit image detection and automating retraining pipelines.

Iterating on LLM apps at scale Learnings from Discord: Ian Webster

Nov 22, 2024 · 18:26

Ian Webster, Senior Staff Engineer at Discord and maintainer of Promptfoo, shares how Discord built and scaled Klyde AI, a chatbot for 200 million users, focusing on evaluation and safety. He argues that evals should be treated as simple, deterministic unit tests that run locally, avoiding complex metrics, and that breaking the system into small, testable pieces (e.g., checking for lowercase output to enforce casual tone) achieves 80% of the goal with 1% of the work. Webster details how Discord mitigated risks like the 'grandma jailbreak' (which originated on Discord) by using an attacker LLM to generate adversarial inputs and a judge to refine them, exposing cracks in safeguards. He advocates for pre-deployment red teaming over live filtering, and describes using Promptfoo for risk assessment across brand, legal, and safety categories. The episode also covers prompt management via Git and Retool, routing with occasional GPT-4 responses to correct model drift, and the challenge of closing the feedback loop due to privacy constraints, relying instead on dogfooding and public examples.

Navigating RAG Optimization with an Evaluation Driven Compass: Atita Arora and Deanna Emery

Nov 12, 2024 · 18:14

Atita Arora (Qdrant) and Deanna Emery (Quotient AI) demonstrate how evaluation-driven optimization improves Retrieval Augmented Generation (RAG) systems, using Qdrant's vector database and Quotient's evaluation platform. They detail a 10-experiment pipeline on Qdrant documentation, starting with naive RAG and incrementally testing chunk sizes, embedding models, LLMs (Mistral, GPT-3.5), re-rankers (MixedBread, Cohere, ColBERT), and hybrid search. Key metrics—faithfulness (hallucination reduction), context relevance, and chunk relevance—guide decisions: increasing chunk size hurt faithfulness (0.76 to 0.72), while smaller chunks with larger retrieval windows improved it. Switching to GPT-3.5 lifted all metrics, but poor context relevance revealed retrieval issues. Adding Cohere re-ranking boosted faithfulness to 0.82, and hybrid search with re-ranking achieved 0.85, proving domain-specific terminology demands tailored search strategies. The episode stresses avoiding over-engineering without metrics, keeping evaluation datasets current, and using domain understanding for systematic RAG improvement.

What It Actually Takes to Deploy GenAI Applications to Enterprises: Arjun Bansal and Trey Doig

Nov 4, 2024 · 21:30

Trey Doig of Echo AI and Arjun Bansal of Log10 recount Echo AI's journey deploying a GenAI-native conversational intelligence platform for billion-dollar retail brands, focusing on the centrality of accuracy. Echo AI ingests all customer conversations, uses LLMs to surface insights at 100% coverage, but must overcome enterprise trust issues by achieving 95% accuracy within seven days. The platform relies on Log10's auto feedback system, which uses AI-based review to match human accuracy with model speed, yielding a 20 F1 point improvement in one use case. The episode details how Echo AI's solution engineers use Log10 to grade summarizations, catch hallucinations, and track model drift, turning human feedback into curated datasets for fine-tuning. Ultimately, the partnership demonstrates a path to self-improving LLM applications through iterative accuracy measurement and improvement.

Productionizing GenAI Models – Lessons from the world's best AI teams: Lukas Biewald

Oct 23, 2024 · 22:36

Lukas Biewald, founder of Weights and Biases, shares lessons from productionizing GenAI models, emphasizing that while AI is easy to demo, it is hard to productionize. He illustrates this with a personal project building a custom Alexa-like device using LLaMA and Whisper, where accuracy improved from 0% to 98% through prompt engineering, switching to Mistral, and fine-tuning with QLoRA. Biewald argues that tracking experiments (including failures) is critical for reproducibility and collaboration, and that a robust evaluation framework—beyond 'testing by vibes'—is essential for iterating and shipping v2. He notes that 70% of the audience had LLM apps in production, yet many lack solid evaluations, and recommends starting with lightweight prototypes and incorporating end-user feedback.

Architecting and Testing Controllable Agents: Lance Martin

Oct 11, 2024 · 2:21:54

Lance Martin presents LangGraph, a graph-based framework for building controllable agents that trade some open-ended flexibility for significantly higher reliability compared to classic React agents, achieving 100% consistent tool-calling trajectories even with an 8B local model. He demonstrates self-corrective RAG patterns like Corrective RAG, SelfRAG, and Adaptive RAG, where the agent grades retrieved documents, checks for hallucinations, and routes to web search when needed. Martin also covers three testing loops: in-app error handling with LangGraph, pre-production evaluation using LangSmith to compare agent answers and tool trajectories against ground-truth datasets, and production monitoring with online evaluators that flag retrieval quality, answer relevance, and hallucinations without reference answers. He shares results from a five-question evaluation showing LangGraph agents achieve 80% answer accuracy and 100% tool-trajectory correctness with Fire Function V2, while React agents with GPT-4o degrade in tool reliability. The talk addresses practical concerns like handling many tools (suggesting RAG for tool selection), multi-turn conversations, and the importance of…

Building and Scaling an AI Agent Swarm of low latency real time voice bots: Damien Murphy

Oct 8, 2024 · 1:07:23

Damien Murphy, Senior Applied Engineer at Deepgram, demonstrates building and scaling low-latency real-time voice bots using Deepgram's new voice agent API, which wraps speech-to-text, LLM, and text-to-speech into a single streaming endpoint. He shows a drive-thru ordering demo with function calling (add/remove items) using GPT-4o, achieving sub-second latency by co-locating components. Murphy explains scaling to millions of concurrent calls through regional Kubernetes clusters and multi-agent swarms (routing, booking, support agents) to reduce complexity and cost. He addresses endpointing challenges, VAD-based barge-in, and cost/quality trade-offs with hosted vs. self-hosted models, noting Deepgram offers 20x cheaper TTS than ElevenLabs and 50ms STT latency self-hosted. The talk emphasizes keeping agents simple, using smallest capable LLMs, and composability for reuse.

From model weights to API endpoint with TensorRT LLM: Philip Kiely and Pankaj Gupta

Sep 13, 2024 · 1:40:01

Philip Kiely and Pankaj Gupta of Baseten lead a workshop on TensorRT-LLM, NVIDIA's high-performance inference framework for LLMs, arguing its use delivers best-in-class throughput and latency on NVIDIA GPUs. They explain that TensorRT-LLM optimizes computational graphs via plugin kernels and in-flight batching, achieving 216 tokens per second and 180ms time to first token on Mistral 7B. The workshop demonstrates building an engine for TinyLlama 1.1B, including FP8 quantization that reduced engine size from 2 GB to 1.2 GB with minimal quality loss. They benchmark a deployed model, showing 7,000 total tokens per second at batch size 64 on an A10G. The presenters compare TensorRT-LLM favorably to VLLM for high-throughput production use and introduce Truss, Baseten's open-source packaging tool, alongside their managed platform for automatic scaling and fast cold starts.

Build enterprise generative AI apps using Llama 3 at 1,000 tokens/s on the SambaNova AI platform

Sep 11, 2024 · 54:34

SambaNova’s Michelle Matern and Petro Milan present their full-stack AI platform, demonstrating how the SN40L RDU chip enables Llama 3 inference at 1,000 tokens per second. They introduce Samba-1, a composition of 92 expert models behind a single endpoint, and benchmark it against GPT-3.5 and GPT-4 on enterprise tasks like information extraction and text-to-SQL. The workshop then builds a RAG-based Q&A system using LangChain, Unstructured, E5-large-v2 embeddings, ChromaDB, and Llama-3-8B-Instruct at 1,000 tokens per second. Attendees set up the environment, load documents, and run inference with real-time metrics showing time to first token of 0.09 seconds and total inference time of 0.65 seconds.

Pydantic is STILL all you need: Jason Liu

Sep 6, 2024 · 15:21

Jason Liu returns to the AI Engineer World's Fair to argue that Pydantic (and his library Instructor) is still all you need for structured output with LLMs. He shows that the core API—response_model, streaming with iterables, and partials for real-time validation—remains unchanged, now supporting Ollama, LlamaCPP, Anthropic, Gemini, and more. Liu demonstrates validators that enforce rules like uppercasing names or verifying receipt totals, reducing errors with automatic retries. He applies structured output to RAG: using a Search model with optional date ranges and source selection, and a Response model with follow-up questions and validated URLs. For extraction, he creates classifiers using Literal types, meeting summaries with action items, and even tables as Pandas DataFrames via custom type hints. Liu’s key takeaway is that one retry often suffices, and as models get faster and smarter, structured output makes LLMs compatible with classical programming—turning generative AI into generating data structures defined by the developer.

Low Level Technicals of LLMs: Daniel Han

Jul 31, 2024 · 2:52:26

Daniel Han of Unsloth explains how to find and fix bugs in open-source LLMs like Gemma, Phi-3, and Llama, covering tokenizer issues, architecture pitfalls, and finetuning optimizations. He details the eight Gemma bugs Unsloth fixed, including a critical RoPE downcasting error that broke positional encoding, and a 2048 sliding window bug in Phi-3. Han walks through transformer internals: attention masking, layer norms, RoPE embeddings, and SwiGLU activation, showing how to derive gradients for custom kernels. He demonstrates Unsloth's 2x faster finetuning with 70% less memory via Triton kernels, and introduces new features: automatic Ollama model file creation, CSV fine-tuning with merged columns, and chunked cross-entropy for large vocabularies. The session includes live Q&A on learning rate schedules, precision trade-offs, and mechanistic interpretability.

Fixing bugs in Gemma, Llama, & Phi 3: Daniel Han

Jul 31, 2024 · 17:42

Daniel Han of Unsloth details eight bugs found in Llama 3, including double BOS tokens, untrained tokens in the base model, and pad tokens equaling EOS tokens, which cause infinite generations. He explains how Unsloth automatically fixes these issues and offers a free Colab notebook for fine-tuning with Ollama. Han also covers tokenization fixes for Gemma and a sliding window bug in Phi-3, emphasizing the importance of correct chat templates and avoiding double BOS tokens. The episode provides concrete code examples and best practices for fine-tuning open-source LLMs to avoid common pitfalls.

Harnessing the Power of LLMs Locally: Mithun Hunsur

Nov 22, 2023 · 17:09

Mithun Hunsur presents llm.rs, a Rust library for running large language models locally, arguing it gives developers ownership, lower latency, and privacy compared to cloud APIs. He explains how quantization makes inference viable on consumer hardware, and shows that llm.rs supports architectures like LLaMA and Falcon through a unified interface. Practical code examples demonstrate customization, and community projects like LocalAI and LLMchain illustrate real-world use. Hunsur shares his own date-extraction pipeline, fine-tuning a small model with GPT-3 data to replace expensive cloud calls. He also cautions about hardware requirements, trade-offs between speed and quality, and ecosystem churn from rapid innovation.

[Workshop] AI Engineering 201: Inference

Nov 7, 2023 · 1:43:16

Charles Frye, instructor of the Full Stack LLM Bootcamp, leads a workshop on AI engineering inference, focusing on the build-versus-buy decision between proprietary and open models. He argues that proprietary models like OpenAI's GPT-4 and Anthropic's Claude are currently more capable but expensive, while open models like LLaMA 2 are less capable but offer hackability, though they may catch up if capabilities requirements saturate. Frye covers inference on end-user devices, noting that running models locally avoids network latency but faces tight memory and power constraints—e.g., a 7B parameter model requires 14 GB, too large for phone RAM. He explains inference-as-a-service (e.g., OpenAI, Replicate) versus self-serving on cloud GPUs or serverless platforms like Modal, highlighting that memory bandwidth is the bottleneck: GPUs have 1.5 TB/s memory bandwidth vs 312 TFLOPS compute, so batching is crucial for throughput. Frye also discusses inference arithmetic, custom silicon like TPUs which offer ~30% better efficiency but not drastic gains, and containerization challenges for GPU workloads.