Agents Everywhere: OpenClaw, Codex, and the Post-Chatbot Shift

Interviewing Alexandre Pereira: 2501.ai and autonomous agents for enterprise IT

In this episode, Bart and Murilo sit down with Alexandre Pereira, co-founder & CEO of 2501.ai — the Paris- and NY-based startup building autonomous AI agents that don't just spot infrastructure incidents, they fix them. We dig into why the real blue ocean is the on-prem, legacy enterprise world the cloud never absorbed (JP Morgan, Goldman Sachs and Société Générale are still mostly on-premise), why roughly half of 2501's work is safety and guardrails rather than the AI itself, and why a privacy-first deployment means they almost never reach for Claude. Alexandre breaks down the model landscape — Qwen vs. GPT-OSS for Windows Server, the “skateboard benchmark”, and why a niche French firewall like Stormshield defeats every frontier model — plus €10M raised across a pan-European cap table and what Europe's pre-seed scene gets right and wrong. 🙋 Guest: Alexandre Pereira — Co-founder & CEO, 2501.ai 🌐 2501.ai [https://www.2501.ai] 📩 Connect with Alexandre on LinkedIn [https://www.linkedin.com/in/pereiraalexandre/] Chapters * (00:00) Intro: autonomous agents that fix enterprise IT * (01:00) Meet Alexandre: from Decathlon at 18 to founding 2501 * (03:50) Genesis: LLM chaining, AutoGPT, and Devin's $100M moment * (10:05) What 2501 does: automating IT maintenance for big corporates * (13:34) Autonomy vs. safety: why half the work is guardrails * (16:57) Earning trust: proving it alongside the customer * (25:00) The on-prem blue ocean nobody is serving * (29:30) Model wars: Qwen, GPT-OSS, and the skateboard benchmark * (35:15) Privacy-first deployments and the future of open weights * (41:03) Roadmap: observability and cybersecurity * (45:30) Raising €10M and Europe's pre-seed reality

Interviewing Stepan from Wyder.io: building Belgium's social coaching app

Great conversation with Stepan Krivosheev, CEO and co-founder of Wyder, the Leuven-born endurance app betting that training should be social and personal at the same time. 🏃 We dig into the origin story, what makes Wyder the "first social coaching app," and how the three first-time founders went full-time just three months ago. Stepan walks us through the marketplace model for finding training partners at your pace, the new AI coach (€14/month, self-hosted on AWS Bedrock for GDPR reasons), and the brutal reality of integrating with Garmin, Coros, Polar and Suunto. We also get into their city-by-city growth playbook (1,400 users in Leuven, now with the move to Antwerp), why micro-influencers care more about brand alignment than money, the willingness-to-pay puzzle among triathletes, and the structural challenge of raising B2C money from European VCs. Plus a fully bootstrapped status update, fresh VLAIO grant in hand, and what the next twelve months and ~30 target cities look like. 🙋‍♂️ Guest: Stepan Krivosheev, CEO & Co-founder, Wyder (Leuven, Belgium) 🔗 Website: https://wyder.io 📩 Connect on LinkedIn: https://www.linkedin.com/in/stepan-krivosheev/ 📸 Instagram: https://www.instagram.com/wyder.app/ 🎵 TikTok: https://www.tiktok.com/@wyderapp * (00:00) - Who Wyder is + meet Stepan * (01:46) - Origin story: from Ironman burnout to a founding team * (06:50) - What makes Wyder stand out + the ideal athlete * (10:00) - How it works: the workout marketplace * (14:18) - Cracking the cold start: 1,400 users, Leuven to Antwerp * (21:40) - Monetization: the €14 AI coach + the B2B lever * (25:27) - Wearables, Garmin's gatekeeping & the data battle * (29:05) - Inside the AI coach: structured vs free-form plans * (38:04) - GDPR, the AI Act & self-hosting the models * (40:08) - Bootstrapped, a VLAIO grant & why EU VCs avoid B2C * (47:45) - The next 12 months + where to find Wyder

Interviewing Marco Ramilli: IdentifAI's de-generative race against deepfakes

In this episode, we sit down with Marco Ramilli, co-founder and CEO of IdentifAI, the Milan-based startup applying a cybersecurity mindset to deepfake detection. Marco, a PhD computer security researcher who worked at NIST and previously founded Yoroi (acquired by Tinexta in 2024), explains how IdentifAI builds "de-generative" AI models that reverse-engineer how generators like Midjourney and Nano Banana work, then catch their outputs in the wild. We get into the verticals where the threats are real today: KYC checks at banks, fraudulent insurance claims (Marco says ~35% of uploaded damage images are now AI-generated), state-sponsored disinformation, journalist fact-checking, and HR processes where attackers join job interviews as deepfake candidates to social-engineer their way into companies. Marco walks through IdentifAI's three products, an API, a web app, and a live "noteseeker"-style agent that silently sits in on video calls and flags fakes in real time, and shares the model accuracy numbers (~98% on internal benchmarks, ~90–92% in the real world) plus what happens when a new generator like Nano Banana drops. We also dig into the €5M Series A led by United Ventures, the EMEA-first go-to-market plan, IdentifAI's Deepfake Intelligence Report (the US accounts for almost half of recorded incidents), the European sovereignty angle, and the EU AI Act. We close with Marco showing us a live, real-time deepfake of his own face, built, he says, with "a couple of Python scripts on your laptop." 🙋‍♂️ Guest: Marco Ramilli, co-founder and CEO, IdentifAI 📩 Connect with Marco on LinkedIn: https://www.linkedin.com/in/marcoramilli [https://www.linkedin.com/in/marcoramilli] 🔗 IdentifAI on LinkedIn: https://www.linkedin.com/company/identifai-labs [https://www.linkedin.com/company/identifai-labs] 🌐 IdentifAI: https://identifai.net [https://identifai.net]

AI’s Collision With Work, Art, and Hardware

Thanks for listening! ❤️ * (09:52) - Yann LeCun Raises $1 Billion to Build AI That Understands the Physical World * (19:44) - Learnings from Paying Artists Royalties for AI-Generated Art * (26:38) - Labor market impacts of AI: A new measure and early evidence * (37:01) - Nvidia prepares AI 'inference' chip launch to counter rising challengers

Agents Everywhere: OpenClaw, Codex, and the Post-Chatbot Shift

We're back! Thanks for listening! ❤️ OpenAI’s acquisition of OpenClaw signals the beginning of the end of the ChatGPT era | 2026-02-17 VentureBeat argues OpenAI’s OpenClaw move is a pivot from chatbots to agents that take actions across apps and systems. The tension is that OpenClaw’s “fast and loose” openness helped it go viral—so what happens when enterprise guardrails and safety expectations move in? https://venturebeat.com/technology/openais-acquisition-of-openclaw-signals-the-beginning-of-the-end-of-the [https://venturebeat.com/technology/openais-acquisition-of-openclaw-signals-the-beginning-of-the-end-of-the?utm_source=tldrai] Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents? (arXiv:2602.11988) | 2026-02-12 A new paper tests whether repo-level context files like AGENTS.md actually help coding agents finish tasks—and finds they can backfire. The punchline is a double hit: lower success rates and more than 20% higher inference cost, hinting that “more context” can mean “more confusion.” https://arxiv.org/abs/2602.11988 [https://arxiv.org/abs/2602.11988] AI Doesn’t Reduce Work—It Intensifies It | 2026-02-09 Simon Willison highlights research suggesting AI can increase work intensity instead of easing it, especially when productivity gains mask burnout. The provocative angle is managerial: if AI boosts throughput, how do organizations prevent that extra capacity from turning into an always-on expectation? https://simonwillison.net/2026/Feb/9/ai-intensifies-work/ [https://simonwillison.net/2026/Feb/9/ai-intensifies-work/] SWE-rebench Leaderboard | n.d. SWE-rebench is trying to solve a messy problem in agent evaluation: benchmarks go stale, and models get “contaminated” by training on the tasks. The leaderboard format makes it feel like a live sport—but the real question is whether continuously refreshed tasks can keep results honest as models ship faster. https://swe-rebench.com/ [https://swe-rebench.com/] Tidbits/extras: Introducing Claude Opus 4.6 | 2026-02-05 Anthropic says Claude Opus 4.6 upgrades its top-tier model for longer, more reliable agentic coding and better performance in large codebases. The headline-grabber is a 1M-token context window in beta—prompting the question of whether bigger memory finally means fewer brittle, lost-in-the-middle failures. https://www.anthropic.com/news/claude-opus-4-6 [https://www.anthropic.com/news/claude-opus-4-6] Introducing GPT-5.3-Codex-Spark | 2026-02-12 OpenAI is pitching GPT-5.3-Codex-Spark as a speed-and-feedback upgrade that makes coding agents feel less like waiting and more like collaborating. The hook is the bet that ultra-fast inference isn’t just convenience—it changes what kinds of multi-step software work people will even attempt with an agent. https://openai.com/index/introducing-gpt-5-3-codex-spark/ [https://openai.com/index/introducing-gpt-5-3-codex-spark/] Introducing GPT-5.3-Codex | 2026-02-05 OpenAI is positioning GPT-5.3-Codex as a coding agent that’s edging toward “do nearly anything on a computer,” not just write snippets. The spicy detail: OpenAI says early versions helped debug and deploy themselves—raising real questions about how fast self-accelerating dev loops can move, safely. https://openai.com/index/introducing-gpt-5-3-codex/ [https://openai.com/index/introducing-gpt-5-3-codex/] OpenClaw creator Peter Steinberger joins OpenAI | 2026-02-15 TechCrunch reports that OpenClaw’s creator, Peter Steinberger, is joining OpenAI as Sam Altman talks up “personal agents” as a core product direction. The interesting tradeoff: Steinberger says he didn’t want to build a standalone company—so can OpenAI keep the project meaningfully open while scaling it? https://techcrunch.com/2026/02/15/openclaw-creator-peter-steinberger-joins-openai/ [https://techcrunch.com/2026/02/15/openclaw-creator-peter-steinberger-joins-openai/] RentAHuman.ai — Hire Humans for AI Agents (MCP Integration) | n.d. RentAHuman.ai is pitching a simple idea: when your agent hits a wall, route the task to a real person instead of failing silently. It frames humans as an on-demand “tool” in the loop—raising a juicy question about where automation ends, and accountability begins. https://rentahuman.ai [https://rentahuman.ai/] NanoClaw — Your personal Claude assistant | n.d. NanoClaw positions itself as a lightweight, local-first Claude assistant: one process, a handful of files, and container isolation for safety. The intriguing wrinkle is its WhatsApp-style interface and per-group memory—suggesting the next wave of “agents” may look more like chatrooms than apps. https://nanoclaw.net/ [https://nanoclaw.net/] * (02:48) - Cambryo “Top of Mind” hiring + AI-native developer workflow (non-article) * (13:01) - OpenAI / OpenClaw “end of ChatGPT era” (VentureBeat) * (13:30) - OpenClaw creator Peter Steinberger joins OpenAI (TechCrunch) * (19:46) - NanoClaw — lightweight/local-first Claude assistant alternative * (21:43) - Introducing Claude Opus 4.6 (Anthropic) * (23:00) - Evaluating **AGENTS.md** repo context files (arXiv:2602.11988) * (33:57) - AI Doesn’t Reduce Work—It Intensifies It (Simon Willison) * (43:50) - SWE-rebench leaderboard + benchmark contamination / freshness * (47:27) - Introducing GPT-5.3-Codex-Spark (OpenAI) * (50:00) - Introducing GPT-5.3-Codex (OpenAI) * (51:19) - RentAHuman.ai — “hire humans for agents” (MCP integration) * (56:36) - Future podcast format + possible rebrand (non-article)