I am super excited to finally announce that we ( @towards_AI ) released our first independent industry-focussed course: From Beginner to Advanced LLM Developer. Put a dozen experts (frustrated ex-PhDs, graduates and industry) and a year of dedicated work, and you get the most practical and in-depth LLM Developer course out there (~90 lessons). It is a one-stop conversion for software developers, machine learning engineers, data scientists, aspiring founders or AI/Computer Science students. We think many millions of LLM Developers will be needed to build reliable customised products on top of foundation LLMs and achieve mass GenAI adoption at companies. We want to help you lead this new field! One of the reasons I quit my PhD in AI (one year 8 days ago) was to build practical solutions that will help others in the real world and improve what exists. Along with our book “Building LLMs for Production,” this course is our attempt to achieve this goal. While I love the world of academia, when I first stepped into the startup world in 2019, it felt like I pretty much knew nothing all over again. I needed to understand the problems in the real-world application of AI and build solutions for it. Not just research but real models, real products, and real people using them. But here’s the thing: grasping these challenges is merely the first step. For the ‘how’ part of it, you need to get into the code, the architecture, the models, the APIs, the deployments, the trials and errors, and the complex and wide varieties of frameworks - you don’t have time to reinvent the wheel in a startup! This discrepancy between academia and the industry is even worse in the LLM era. So we’ve gathered everything we worked on building products and AI systems and put them into one super practical industry-focused course. Right now, this means working with Python, OpenAI, Llama 3, Gemini, Perplexity, LlamaIndex, Gradio, and many other amazing tools (we are unaffiliated and will introduce all the best LLM tool options). It also means learning many new non technical skills and habits unique to the world of LLMs. Even though the course is super practical (oriented towards building a real-world project), we believe the course teaches concepts that will stay relevant for a long time even as LLMs get better, such as reducing hallucinations, customising to specific companies and tasks, teaching how to work with them, some cool theory, practical tips and more. The only skill required for the course is some Python (or programming) knowledge. We cover the full stack of learning to build on top of foundation LLMs - from choosing a suitable LLM application to collecting data, iterating on many advanced techniques (RAG, fine-tuning, agents, audio, caching), integrating industry expertise and deploying. Our students will create a working product, which we certify, and we also provide instructor support in our dedicated Discord channel. This could become the seed of a startup, a new tool at your company or portfolio project for your LLM job applications. You can find all the lesson titles and more information on the course page (or DM me). I also want to thank some of our amazing team at Towards AI that helped achieve this, including @omar_solano1 , Fabio , Rucha , @_LouiePeters , Arjun, Jaiganesan and everyone in the team that contributed to this course. A big thanks to some of our reviewers and my friends for the support, @daansan_ml , @SerranoAcademy , @rohanpaul_ai , @tunguz , @antgrasso , @ceobillionaire , @Scobleizer , @swyx , @hasantoxr and many others... “From Beginner to Advanced LLM Developer” is now available on the Towards AI Academy: tinyurl.com/5n9y32y4 #llms #llm #llmops #llmdeveloper

The US government just made Anthropic suspend two of its own models. Fable 5 and Mythos 5, gone for everyone, everywhere. Just three days after launch. The whole thing started over a "jailbreak." But when Anthropic looked at it, the jailbreak was basically asking the model to read code and fix bugs. Something devs do every day. Anthropic is calling it a misunderstanding and complying for now. The government is treating these models as too risky to leave open.

The 8-step order I use for shipping AI agents in 2026: 1. Filter noisy tool outputs 2. Load tools only when needed 3. Clean cached history before reusing it 4. Compress long logs and terminal outputs 5. Store memory outside the context window 6. Compact manually around 40% 7. Add retrieval behind the system 8. Keep autocompact as the last resort What I like about this order is that each step reduces pressure on the context window before the next layer gets added. The result is not just lower token usage. The agent stays much more stable during long sessions and keeps track of what it’s actually doing. That’s becoming a very important skill in agent engineering now. Not just building agents. Managing their context properly.

Stop prompting your coding agents. The people building Claude Code and Codex don't prompt anymore — they design loops. New video: what loop engineering actually is, the 5 building blocks, and the 2 traps that burn your token budget. youtu.be/NjXIIH9vcv0

Claude Fable 5 might be a warning sign for AI research. Not because of safeguards. Because of invisible safeguards. Anthropic's system card confirms that Fable 5 deliberately limits its effectiveness on requests targeting frontier LLM development — pretraining pipelines, distributed training infrastructure, ML accelerator design. The model still answers. But the answer can be quietly degraded through prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT). And unlike the cybersecurity and biology safeguards, which visibly reroute you to Opus 4.8, Anthropic states these ones "will not be visible to the user." That part is the problem. I understand the safety concern. I understand not wanting frontier models to accelerate unsafe actors and competitors. Anthropic even notes that using Claude to build competing models already violates its Terms of Service. But a refusal is transparent. Silent degradation is not. If you work on training infrastructure or efficiency research, you can no longer fully trust a weak answer: is it the model's limit, or an intervention? Anthropic itself says it is still improving the precision of its detection — meaning false positives are part of the deal. A silent handicap is an unlogged confounder in every ML engineering workflow built on top of it. And honestly, it is disappointing to see this come from Anthropic first. I would have expected this kind of ladder-pulling from OpenAI. But it seems increasingly likely that all frontier labs will converge on the same playbook: 1. Keep the unrestricted model (Mythos 5) for approved partners. 2. Ship a constrained version (Fable 5) to everyone else. 3. Hide the constraints behind "safety." Anthropic estimates ~0.03% of traffic is affected, concentrated in fewer than 0.1% of organizations. The number is small. The precedent is not. This is bad for open research, bad for small labs, and bad for trust. Safety matters. But users should know when the tool is intentionally limiting them. What else are they steering that we don't know of? Where's the limit? What do you think: is invisible capability degradation ever acceptable? #AI #LLMs #Anthropic #OpenResearch #AISafety

I realized something a bit stupid about myself during my 2-week vacation visiting Switzerland. Even on vacation, my brain kept trying to optimize everything. Can I answer a few emails quickly? Maybe work a bit from the coffee shop. Maybe wake up early and squeeze in a run before the day starts. At some point I noticed I wasn't even relaxing anymore. I was just turning “rest” into another thing to optimize. The weird part is that I genuinely love what I do. So it never feels bad in the moment. But after a few years of working like this, your body starts reminding you that stress still exists even if your brain ignores it. That’s probably the hardest part when your work is also your hobby. You stop noticing when you actually need a break. So during this trip, I tried something very rare for me: being fully offline and actually present. Way harder than I expected 😅

mythos will be bad ON PURPOSE on ai "frontier llm research" tasks, this is very very sad for the research community also the fact that this is un purpose not visible to the user is crazy

New video ! 🚀 How AI Image Generation Actually Works How image models actually understand text, photos, and edits... youtu.be/DCOXcnOntmg

2 weeks ago, I was in Switzerland helping run a 7-hour AI engineering workshop. I almost didn’t attend, but I’m glad I did. Here’s what surprised me most: The people weren’t just interested in AI… The workshop had around 34 attendees across every experience level imaginable: People who had never coded before Software architects • Product managers CEOs and executives Engineers in their 20s People in their 60s and 70s It was one of the most engaged audiences I’ve ever seen. I planned hours of material around agentic engineering and AI coding workflows… But a massive part of the workshop became discussion. People constantly interrupted with thoughtful questions: "How do these systems fail in production?" "How do you evaluate agentic coding?" "What happens when agents become unreliable?" "Where does MCP fit?" "How should engineering teams adapt?" The biggest takeaway for me: AI engineering is becoming cross-functional extremely fast. It’s no longer just engineers trying to understand these systems. The second thing I realized: The best conversations happen when you stop thinking hierarchically. During the conference, I had the chance to speak with researchers, principal architects, founders, and engineers from companies like: Google DeepMind Hugging Face Cisco And what stood out is that the strongest people rarely act important. Treating people as equals completely changes the quality of interactions. Third lesson: Networking works much better when you stop trying to maximize every interaction. You don’t need to force every conversation into an opportunity. Sometimes: Introducing yourself Talking for a few minutes Moving on is enough Overall, this trip reminded me how global AI engineering has become. As someone coming from Romania, traveling to Switzerland to teach AI systems still feels surreal sometimes. But it also reinforced good ideas travel very fast in this industry. Massive thanks to Louis-Francois Bouchard (@Whats_AI) and the @towards_AI team for inviting me and making the whole experience happen.

Everyone thinks OpenAI just signed the exact deal Anthropic refused. That's not the full picture. Here's what actually happened. OpenAI got the deal first, then a few days later went back and quietly tightened their own wording. The update said their tools wouldn't be used for domestic surveillance of US persons, including through commercially bought personal data, and that certain intelligence uses would need a fresh agreement. They narrowed the autonomous weapons language too. So OpenAI didn't say yes to everything Anthropic said no to. Anthropic just drew the line earlier, harder, and out in the open. OpenAI found a path through to get the deal, then moved the terms closer to where Anthropic had been pushing all along. That's the part people miss. It was never one clean company saying no and another saying yes. Both were already deep inside the government, and both ended up near the same place. They just got there very differently.