About
Hi, I am Louis-François (or just Louis, loo·ee, French pronunciation), from Montreal…
Services
Articles by Louis-François
-
How to Get an AI Engineer Job in 2026
How to Get an AI Engineer Job in 2026
Good morning, AI enthusiasts! AI engineering interviews are still being described like software interviews with a few…
-
How I Actually use Coding Agents
How I Actually use Coding Agents
Good morning, AI enthusiasts! Coding agents have become good enough that the tool choice barely matters for many tasks.…
2 Comments -
Why We Stopped Compacting Our Agent's Context
Why We Stopped Compacting Our Agent's Context
Good morning, AI enthusiasts! If you’ve missed it, the recording of our live workshop at AI Engineer World’s fair is…
2 Comments -
We're Launching the Towards AI Mentorship
We're Launching the Towards AI Mentorship
Good morning! I am genuinely excited to finally share this. When companies bring our team into an AI project, the…
-
How I'd Learn AI Engineering in 2026
How I'd Learn AI Engineering in 2026
Good morning! A working AI app can now be a terrible learning signal. Codex, Claude Code, and Cursor can put together a…
2 Comments -
How I Automated 90% of My YouTube Workflow
How I Automated 90% of My YouTube Workflow
Good morning! I used to spend two to three days researching, writing, and packaging one YouTube video. Every single…
-
Graph Engineering Explained: What Actually Changed
Graph Engineering Explained: What Actually Changed
Good morning! Apparently, loop engineering died before some of you even had time to try it. Peter Steinberger posted…
5 Comments -
My AI Engineer keynote is live today
My AI Engineer keynote is live today
Good morning! Small exciting update today: our AI Engineer World’s Fair online talk was selected as a keynote for the…
2 Comments -
Why Microsoft Trained MAI-Thinking-1 Without Synthetic Data
Why Microsoft Trained MAI-Thinking-1 Without Synthetic Data
Good morning! Most AI model releases try to win you over with benchmarks. This one is more interesting because of what…
-
The Hidden Cost of Cheap AI APIs
The Hidden Cost of Cheap AI APIs
Good morning! Last week, I discovered something quite alarming. You can get the GPT or Claude API 90% cheaper.
2 Comments
Activity
47K followers
-
Louis-François Bouchard shared thisIf you know anyone that is down for a coffee, or even interview, please tag them!
-
Louis-François Bouchard shared thisMy co-founder Louie Peters and I will be in San Francisco from September 12–18. Far away from London and Montreal! We’d love to meet some founders, CTOs, FDEs, and other people working closely with AI. We’re also looking for a few people who’d be open to doing a short, informal interview with us while we’re there. We want to hear what you’re building, what you’re struggling with, and what you think is still missing when it comes to AI inside companies. If you’re in SF that week and this sounds interesting, send me a message.
-
Louis-François Bouchard shared thisWe tested 146 model variants on 10 script-writing tasks, 5 runs each, scored by a 3-judge panel on tone, voice, and whether a human would actually like it enough to read it out loud without cringing. Our goal is to measure creative writing quality. Not agentic or coding performance, or even one-shot generation performance. We provide a very detailed prompt with all the info the model needs to do its job (in theory). Here are some insights from the main family of models with GPT, Fable, Gemini, Muse... (see image for more info): 1. If you are using sol 5.6, skip the ultra setting. It doubles the bill from $0.21 to $0.43 per task and buys only 66 Elo. 2. Extra thinking is worth paying for on the small models, not the big ones. Luna gains 230 Elo for 3× the cost. Sol gains 66 for 2×. Grok 4.6 and Gemini 3.1 Pro get worse when you turn it up. 3. Claude Opus 5 is the value pick of the whole chart. Max effort is $0.42 a task, and it out-writes every GPT config on the board — including GPT-6 Astra, which costs 8× more. I know this one is controversial, but I also see it using it personally. And that's what you should trust in any case: you own usage, not a benchmark. 4. Never pay for Sonnet 5's max effort. It burns $2.96 a task, 7× what Opus 5 max costs, and writes 216 Elo worse. 5. GPT-6 Astra is the most expensive way to be mid. $3.42 a task at max effort, 246 Elo behind Opus 5 max. 6. Meta Muse Spark 1.3 is the budget outlier: $0.034 a task, and it out-writes every Gemini 3.1 Pro config — which costs 4× more. 7. Claude Fable 5 still tops the board. $4.36 a task at max effort: 10× Opus 5 max for 72 Elo. Pay it only if the words are the product or you still have your subscription tokens. 8. Claude Fable 5.1 is not a writing upgrade. At max effort it costs 25% more than Fable 5 ($3.15 vs $2.52) and scores 6 Elo lower, which is inside the noise band.
-
Louis-François Bouchard shared thisSuper excited for tomorrow! Join in :)Louis-François Bouchard shared this🤖 Upcoming free webinar about personal AI agents. I already have my own coding agent and it's become a valuable part of my workflow. But I've learned that the fastest way to improve is often not by building more, but by learning from how others build. That's why I'm looking forward to an upcoming webinar about AI engineering systems and agent workflows by Louis-François Bouchard on Eventbrite . https://lnkd.in/ewjAhFuj What interests me most isn't a new tool or a new model. It's the real-world use cases. 💡 How are people structuring their agents? 💡 What workflows are saving them the most time? 💡 How do they manage context, memory, and feedback loops? 💡 What surprising use cases have they discovered that I haven't thought of yet? I'm hoping to leave this session with a few practical ideas, some fresh inspiration, and maybe a challenge to rethink parts of my current setup. 🚀 The AI space is moving incredibly fast, but sharing experiences is still one of the best ways to learn. For those of you working with coding agents or AI assistants: 👉 What's the most valuable task you've delegated to an agent so far? #AI #AgenticAI #AIEngineering #Eventbrite #LearningInPublic
-
Louis-François Bouchard shared thisCost is now part of the AI engineering interview. If you're using a heavy reasoning model for something that DeepSeek-v4 Flash could handle, I'm going to ask why. The more AI gets built into products, the more these choices matter. A model can work perfectly and still be the wrong choice if the cost doesn't make sense. A few years ago, model choice was mostly a technical decision. Now engineers have to think about the architecture and the bill that comes with it too. That’s probably one of the big changes I’ve seen in the AI engineer role.
-
Louis-François Bouchard reposted thisBig news from our internal writing benchmark (early results): GPT-6 ... is surprisingly disappointing I definitely did not expect that... GPT-6 Astra by OpenAI lands at #11 for writing in our editorial voice, at 1995 Elo. That is below its predecessor. GPT-5.6 Sol sits #6 at 2156. It also runs at about $0.26 per script, roughly 1.8x GPT-5.6. 𝗪𝗵𝗲𝗿𝗲 𝗚𝗣𝗧-𝟲 𝗱𝗼𝗲𝘀 𝘄𝗶𝗻: (𝗮𝗻𝘁𝗶-)𝘀𝗹𝗼𝗽. 90.9, first on the board, ahead of GPT-5.6 Sol at 90.1 and Claude Fable 5.1 at 90.0. It uses fewer stock AI phrases and handles numbers more cleanly than anything else we have tested. If your problem is copy that reads like a bot wrote it, this is the best fix we have measured. 𝗜𝘁'𝘀 𝗮𝗹𝘀𝗼 𝘁𝗵𝗲 𝗰𝘂𝗿𝗿𝗲𝗻𝘁 𝗯𝗲𝘀𝘁 𝗺𝗼𝗱𝗲𝗹 𝗮𝘁 𝗳𝗼𝗹𝗹𝗼𝘄𝗶𝗻𝗴 𝘁𝗵𝗲 𝗿𝗲𝗾𝘂𝗲𝘀𝘁𝗲𝗱 𝗹𝗲𝗻𝗴𝘁𝗵, like "I want a 12-minute speaking time script". 95.2. Claude Fable 5.1 is next at 94.8. Only for Max effort!! At the API default effort that number drops to 83.7. 𝗪𝗵𝗲𝗿𝗲 𝗶𝘁 𝗹𝗼𝘀𝗲𝘀 𝘁𝗵𝗲 𝗺𝗼𝘀𝘁 𝗽𝗼𝗶𝗻𝘁𝘀: 𝘁𝗼𝗻𝗲 𝗮𝗻𝗱 𝘃𝗼𝗶𝗰𝗲. 83.6, #55 of 146 configurations, against a leader at 89.4. 𝗧𝗵𝗲 𝗲𝗳𝗳𝗼𝗿𝘁 𝗹𝗮𝗱𝗱𝗲𝗿 𝘄𝗼𝗿𝗸𝘀 𝗼𝘂𝘁 𝗾𝘂𝗶𝘁𝗲 𝘄𝗲𝗹𝗹 𝗵𝗲𝗿𝗲: default effort scores 1848 Elo (#41 of 146) in 85 seconds per script. Max effort scores 1995 (#30) in 526 seconds. A bit more expensive, but since I run it through a subscription right now, it's quite negligible. You mostly pay with time. 𝗣𝗿𝗮𝗰𝘁𝗶𝗰𝗮𝗹 𝘁𝗮𝗸𝗲𝗮𝘄𝗮𝘆 𝗳𝗿𝗼𝗺 𝘁𝗵𝗲𝘀𝗲 𝗿𝗲𝘀𝘂𝗹𝘁𝘀: If you draft in your own voice with GPT, stay on GPT-5.6 Sol for now. Ultra effort sits #6 at 2156 with better tone, at almost half the price and less than half the wait. But I'd switch to Fable or even Kimi for the best results. GPT-6 seems to struggle with subjective instructions. Reach for GPT-6 Astra when the brief is numbers-heavy and length-exact and a human rewrites the voice anyway. It seems to produce great, accurate, in-depth results that follow quantifiable instructions.
-
Louis-François Bouchard shared thisBig news from our internal writing benchmark (early results): GPT-6 ... is surprisingly disappointing I definitely did not expect that... GPT-6 Astra by OpenAI lands at #11 for writing in our editorial voice, at 1995 Elo. That is below its predecessor. GPT-5.6 Sol sits #6 at 2156. It also runs at about $0.26 per script, roughly 1.8x GPT-5.6. 𝗪𝗵𝗲𝗿𝗲 𝗚𝗣𝗧-𝟲 𝗱𝗼𝗲𝘀 𝘄𝗶𝗻: (𝗮𝗻𝘁𝗶-)𝘀𝗹𝗼𝗽. 90.9, first on the board, ahead of GPT-5.6 Sol at 90.1 and Claude Fable 5.1 at 90.0. It uses fewer stock AI phrases and handles numbers more cleanly than anything else we have tested. If your problem is copy that reads like a bot wrote it, this is the best fix we have measured. 𝗜𝘁'𝘀 𝗮𝗹𝘀𝗼 𝘁𝗵𝗲 𝗰𝘂𝗿𝗿𝗲𝗻𝘁 𝗯𝗲𝘀𝘁 𝗺𝗼𝗱𝗲𝗹 𝗮𝘁 𝗳𝗼𝗹𝗹𝗼𝘄𝗶𝗻𝗴 𝘁𝗵𝗲 𝗿𝗲𝗾𝘂𝗲𝘀𝘁𝗲𝗱 𝗹𝗲𝗻𝗴𝘁𝗵, like "I want a 12-minute speaking time script". 95.2. Claude Fable 5.1 is next at 94.8. Only for Max effort!! At the API default effort that number drops to 83.7. 𝗪𝗵𝗲𝗿𝗲 𝗶𝘁 𝗹𝗼𝘀𝗲𝘀 𝘁𝗵𝗲 𝗺𝗼𝘀𝘁 𝗽𝗼𝗶𝗻𝘁𝘀: 𝘁𝗼𝗻𝗲 𝗮𝗻𝗱 𝘃𝗼𝗶𝗰𝗲. 83.6, #55 of 146 configurations, against a leader at 89.4. 𝗧𝗵𝗲 𝗲𝗳𝗳𝗼𝗿𝘁 𝗹𝗮𝗱𝗱𝗲𝗿 𝘄𝗼𝗿𝗸𝘀 𝗼𝘂𝘁 𝗾𝘂𝗶𝘁𝗲 𝘄𝗲𝗹𝗹 𝗵𝗲𝗿𝗲: default effort scores 1848 Elo (#41 of 146) in 85 seconds per script. Max effort scores 1995 (#30) in 526 seconds. A bit more expensive, but since I run it through a subscription right now, it's quite negligible. You mostly pay with time. 𝗣𝗿𝗮𝗰𝘁𝗶𝗰𝗮𝗹 𝘁𝗮𝗸𝗲𝗮𝘄𝗮𝘆 𝗳𝗿𝗼𝗺 𝘁𝗵𝗲𝘀𝗲 𝗿𝗲𝘀𝘂𝗹𝘁𝘀: If you draft in your own voice with GPT, stay on GPT-5.6 Sol for now. Ultra effort sits #6 at 2156 with better tone, at almost half the price and less than half the wait. But I'd switch to Fable or even Kimi for the best results. GPT-6 seems to struggle with subjective instructions. Reach for GPT-6 Astra when the brief is numbers-heavy and length-exact and a human rewrites the voice anyway. It seems to produce great, accurate, in-depth results that follow quantifiable instructions.
-
Louis-François Bouchard reposted thisI searched my name on Google today. And this came up: “Demystifying the AI black box.” I actually like that. Because that’s pretty much what I’ve been trying to do with my YouTube channel. Every week there’s something new in AI, and it’s easy to feel like you’re already behind. I like taking those things apart and figuring out what’s actually worth knowing. Then explaining it without making it sound like a research paper.
-
Louis-François Bouchard reacted on thisLouis-François Bouchard reacted on thisI benchmarked every MTP depth from 0 to 15 on a Qwen3.8-27B running on an RTX 4090. Because apparently “try 1–6” wasn’t enough. 😅 The result: ~2× faster decoding, with `draft_num_predict=5` coming out on top. But the interesting part is what happens after 5: eventually, adding more speculative tokens doesn’t make things faster (it makes them dramatically slower). Up to 5.3× slower than disabling MTP entirely. And the cutoff changes depending on the quantization. If you’re running local LLMs, there’s a useful lesson here: “use a low value” is not really a configuration strategy. I wrote up the full benchmark, methodology, numbers, and what I’d actually recommend using. 👉 Full article at Towards AI, Inc. : https://lnkd.in/eQDypbdd My 4090 sacrificed an evening so you don't have to. Louis-François Bouchard. Louie Peters. #Ollama #LLM #AI #Qwen #ArtificialIntelligence #AICodingAgent #CodingAgentI Benchmarked Every MTP Depth on Qwen3.8–27B-UD-Q4. Here You Are.I Benchmarked Every MTP Depth on Qwen3.8–27B-UD-Q4. Here You Are.
-
Louis-François Bouchard reacted on thisLouis-François Bouchard reacted on thisI’m happy to share that I’ve completed the Agentic AI Engineering certification from Towards AI. The course focused heavily on building agentic systems beyond the prototype stage where reliability, security, observability, and evaluation become engineering concerns. As part of the certification, I built a financial transaction resolution MCP server. The system takes a customer’s question about a credit card transaction, uses tools and resources to gather the relevant evidence, explains what it finds, and creates a case only after the customer explicitly confirms they want to proceed. Check out the project and demo: https://lnkd.in/gPb8KT4M A few things I took away from the course: Smart context management plays a critical role in the success of agentic systems. I learned how to manage dynamic memory so the agent has exactly what it needs to solve a complex task without getting lost in the noise of a massive context window. I spent time thinking about what should be part of the current state, what should persist, and what the agent simply doesn’t need. When something goes wrong, the final response alone doesn’t tell you much. I found tracing the LLM calls, tool calls, state transitions, latency, and token usage much more useful for understanding what the system actually did and where it failed. A response can look correct even when the agent took the wrong path to get there. My evals therefore included questions such as: Did it use the right evidence? Did it call the appropriate tool to resolve the customer query? Getting an agent to work in a demo is one problem. Making its behavior traceable, testable, and controlled when it interacts with real systems is a very different engineering problem. Huge thanks to Towards AI, Inc., Louis-François Bouchard, and @Paul lusztin for putting together this course. If you want to learn how to build production level agentic systems, check out the course link in the comments.
-
Louis-François Bouchard liked thisLouis-François Bouchard liked thisExcited to share I’ll be flying to SF tomorrow for Anthropic’s Claude SMB Trainer Workshop! All the SMBs I work with are already using Claude, yet are struggling to get the most out of it. The analogy a friend recently told me is that Anthropic is selling businesses the keys to a F1 racer, but not teaching them how to use it. That’s the gap that I (and others at the workshop) help close. This is also the 1st step of receiving an “Authorized Training Partner” badge from Anthropic :) I'll have some free time on the trip, so if you have any SF recs, drop them in the comments 👇 P.S. I love bread
Experience
Education
Publications
-
Building LLMs for Production
Towards AI
“This is the most comprehensive textbook to date on building LLM applications - all essential topics in an AI Engineer's toolkit."
- Jerry Liu, Co-founder and CEO of LlamaIndexOther authorsSee publication -
Ignore This Title and HackAPrompt: Exposing Systemic Vulnerabilities of LLMs through a Global Scale Prompt Hacking Competition
The 2023 Conference on Empirical Methods in Natural Language Processing
Large Language Models (LLMs) are deployed in interactive contexts with direct user engagement, such as chatbots and writing assistants. These deployments are vulnerable to prompt injection and jailbreaking (collectively, prompt hacking), in which models are manipulated to ignore their original instructions and follow potentially malicious ones. Although widely acknowledged as a significant security threat, there is a dearth of large-scale resources and quantitative studies on prompt hacking. To…
Large Language Models (LLMs) are deployed in interactive contexts with direct user engagement, such as chatbots and writing assistants. These deployments are vulnerable to prompt injection and jailbreaking (collectively, prompt hacking), in which models are manipulated to ignore their original instructions and follow potentially malicious ones. Although widely acknowledged as a significant security threat, there is a dearth of large-scale resources and quantitative studies on prompt hacking. To address this lacuna, we launch a global prompt hacking competition, which allows for free-form human input attacks. We elicit 600K+ adversarial prompts against three state-of-the-art LLMs. We describe the dataset, which empirically verifies that current LLMs can indeed be manipulated via prompt hacking. We also present a comprehensive taxonomical ontology of the types of adversarial prompts.
Other authorsSee publication -
Artificial Intelligence clearly explained
YouTube
See publicationArtificial Intelligence clearly explained
Projects
-
Start Machine Learning in 2021 - Become an expert for free!
- Present
See projectA complete guide on how to start learning machine learning (ML), artificial intelligence (AI) in 2021 without ANY background in the field.
-
Explaining the most used AI terms in a minute
- Present
See projectExplaining the most used AI terms in a minute.
Honors & Awards
-
50 000 Subscribers on YouTube: What's AI
YouTube
-
Canada Graduate Scholarships – PhD program (NSERC)
NSERC
-
Featured in Onalytica's Who's Who in AI? Report
Onalytica
https://onalytica.com/blog/posts/interview-with-louis-bouchard/
-
Canada Graduate Scholarships – Master’s program (NSERC)
NSERC
-
10 000 Subscribers on YouTube: What's AI
YouTube
Reached 10 000 Subscribers on YouTube!
-
NSERC Undergraduate Student Research Awards
NSERC
Received the NSERC undergrad award for my final year’s project in computer vision.
-
J.A. DeSève scholarship
J.A. DeSève
Received J.A. DeSève scholarship for having kept a good average and perseverance in the baccalaureate.
Languages
-
French
Native or bilingual proficiency
-
English
Native or bilingual proficiency
Other similar profiles
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More