Bluebird

As Managing Partner IT at Bluebird Recruitment, I am responsible for building out our IT recruitment team with ex-industry experts and connecting top IT talent with leading companies throughout Europe. 

We focus on various IT areas, from software development, cloud & infrastructure, data science & engineering, product roles and more. 

Determined to transform the recruitment space, my goal is to integrate authentic industry expertise, fundamentally changing the sector. By utilising my technical background and understanding of the industry, I am committed to making an impact.

Managing Partner IT

ASML

At ASML, a leader in the semiconductor industry, I contributed to developing a platform designed to provide insights into machine operations. This platform played a crucial role in monitoring for errors, facilitating timely fixes, and employing artificial intelligence to anticipate potential failures.

Due to the proprietary and sensitive nature of the technology and processes at ASML, much of the detailed information about my role and the project specifics is classified. This experience has honed my skills in handling confidential data and working within strict operational security guidelines, reinforcing the importance of discretion and integrity in high-stakes environments.

𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝘆: 𝗦𝗲𝗺𝗶𝗰𝗼𝗻𝗱𝘂𝗰𝘁𝗼𝗿

𝗧𝗲𝗰𝗵 𝘂𝘀𝗲𝗱: 𝗝𝗮𝘃𝗮, 𝗔𝗻𝗴𝘂𝗹𝗮𝗿, 𝗔𝗶𝗿𝗳𝗹𝗼𝘄 𝗳𝗼𝗿 𝗗𝗮𝘁𝗮 𝗘𝗻𝗴.

Senior Fullstack Engineer

Alphabet Nederland

At Alphabet Lease (part of BMW Group), a leasing company, my role involved developing modular components for a complex system and leading several frontend projects. As a Chapter Lead I spearheaded technical innovation within our eco-system. This included conducting Angular upgrades and enhancing the codebase for better performance. The frontend team, comprising 10 members, was responsible for overseeing and managing 20 separate Angular projects, each an individual application.

I also contributed to the refactoring of a key application, transitioning it to a monorepo structure with Nx to improve maintainability and scalability. 

𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝘆: 𝗔𝘂𝘁𝗼𝗺𝗼𝘁𝗶𝘃𝗲

𝗧𝗲𝗰𝗵 𝘂𝘀𝗲𝗱: 𝗔𝗻𝗴𝘂𝗹𝗮𝗿, 𝗝𝗮𝘃𝗮, 𝗞𝘂𝗯𝗲𝗿𝗻𝗲𝘁𝗲𝘀

Frontend Chapter Lead

Lyfo

At Lyfo, a provider of advanced mobile coverage solutions critical for ensuring reliable communication, I developed an Angular application made for configuring the backend for their clients, including the Dutch police. The code that was written is quality tested by an external party and was found to be excellent.

𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝘆: 𝗧𝗲𝗹𝗲𝗰𝗼𝗺 (𝘀𝗼𝗳𝘁𝘄𝗮𝗿𝗲)

𝗧𝗲𝗰𝗵 𝘂𝘀𝗲𝗱: 𝗔𝗻𝗴𝘂𝗹𝗮𝗿, 𝗚𝗼, 𝗞𝘂𝗯𝗲𝗿𝗻𝗲𝘁𝗲𝘀

Lead Frontend Engineer

Datlinq

At DatLinq, a Dutch company known for its data solutions in the foodservice industry, I worked on the SaaS platform (DOX). My responsibilities included the development and enhancement of Data Outlet X and Salesmapp API. This platform gave our clients in the foodservice market data-driven decision making capabilities.

𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝘆: 𝗦𝗮𝗮𝗦 & 𝗙𝗼𝗼𝗱𝗦𝗲𝗿𝘃𝗶𝗰𝗲

𝗧𝗲𝗰𝗵 𝗨𝘀𝗲𝗱: 𝗔𝗻𝗴𝘂𝗹𝗮𝗿 𝟴+, 𝗝𝗮𝘃𝗮, 𝗚𝗼𝗟𝗮𝗻𝗴, 𝗡𝗼𝗱𝗲, 𝗚𝗿𝗮𝗽𝗵𝗤𝗟

Lead Fullstack Engineer

Ministerie van Volksgezondheid, Welzijn en Sport

Increasing mobility of citizens between EU countries results in a growing number of people seeking healthcare outside their own region or country. To facilitate the exchange of patient data among countries, the European Union collaborates with member states on an international infrastructure known as eHealth Digital Services & Infrastructure (eHDSI). The Netherlands is integrating into this system of national contact points with its own national contact point called the National Contact Point for eHealth (NCPeH). Through the NCPeH, healthcare professionals can digitally retrieve patient summaries from patients in EU countries requiring unplanned care in the Netherlands.

𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝘆: 𝗚𝗼𝘃𝗧𝗲𝗰𝗵

𝗧𝗲𝗰𝗵 𝘂𝘀𝗲𝗱: 𝗔𝗻𝗴𝘂𝗹𝗮𝗿, 𝗝𝗮𝘃𝗮, 𝗞𝘂𝗯𝗲𝗿𝗻𝗲𝘁𝗲𝘀, 𝗢𝗽𝗲𝗻𝗡𝗖𝗣.

codecentric Nederland B.V.

Codecentric is a consultancy firm. Since the start I have worked on one project:
Project: MOJO (Read below)

Fullstack Engineer

Mojo Concerts

At MOJO Concerts, a leading promoter of live music and entertainment events, I was tasked with developing an application named UNA. UNA is designed to streamline event access control, scheduling, and workflow management, playing a critical role in enhancing the operational efficiency of live events. This application utilizes a modern tech stack and leverages cloud-based solutions to ensure scalability and reliability.

𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝘆: 𝗦𝗮𝗮𝗦 𝗳𝗼𝗿 𝗮𝗰𝗰𝗲𝘀𝘀 𝘁𝗼 𝗲𝘃𝗲𝗻𝘁𝘀 & 𝗰𝗼𝗻𝗰𝗲𝗿𝘁𝘀

𝗧𝗲𝗰𝗵 𝘂𝘀𝗲𝗱: 𝗔𝗻𝗴𝘂𝗹𝗮𝗿, 𝗡𝗼𝗱𝗲, 𝗚𝗿𝗮𝗽𝗵𝗤𝗟, 𝗖𝘆𝗽𝗿𝗲𝘀𝘀, 𝗚𝗖𝗣, 𝗔𝗪𝗦

Working while being a student

During my studies, to gain practical experience, I wanted to work. I faced the challenge that no company was willing to hire me part-time. Driven by resilience and determination, I didn't let this stop me. Instead, I chose to start my own freelancing 'company' as a way to gain that experience. I've worked on a variety of projects, ranging from small websites for local artists to large-scale applications like a planning system for a major transportation company. This experience has given me a versatile skill set, making me broadly deployable in various IT roles.

The tech stack I've used includes: WordPress, PHP, Laravel, JavaScript, AngularJS, Angular, React, HTML/CSS, and React Native.

As a former developer, I redefine the kind of developer you think you’re looking for and see beyond the resume to ensure you hire the real deal.

If you're a developer or hiring manager, feel free to send me a message on work related things.

If you're a human being you can always send me a message with anything random because I like engaging with you nonetheless 😉

🚀 Dennis van Valderen

Brenn Hill

If you want to use GenAI and LLMs to double (or more!) the output of your engineering team - safely, reliably, verifiably - talk to me.  I've successfully implemented GenAI development across different industries and teams and can *guarantee* an 2x (being very conservative) improvement in your company's velocity.  

How many people can make that claim? Not many.

I'm a global multi-industry tech executive with 17 years experience building and leading engineering teams and products with international scale across USA, Europe, and Asia.  I focus on driving business impact with technology instead of tech for it's own sake, separating what works from what is just more hype. 

Author of Gasoline - Fuel for the AI Fire (https://github.com/brennhill/gasoline)

- Define and execute product development and technical strategies to increase revenue, decrease costs, and apply innovation, AI, and machine learning to modern business problems.

- Experience from small startup up through publicly traded global engineering teams. No matter where you are on your growth trajectory as a company, I have helped get teams to the next level.  

- Thought leader and speaker at conferences and host tech industry events in the USA, Germany, and Asia.

More about me:
🌍 Proven success leading teams and launching products in the USA, Europe, and Asia
🚀 Built and scaled teams up to 75 engineers; led platform modernization, cloud migration, AI and ML integration, and agile transformation
🤝 Mentor to engineering leaders at Atlassian, Delivery Hero S.E, VNG, Taxfix, and more  
🛠️ Hands-on with large scale distributed system design and a broad tech stack: AWS, GCP, Java/Kotlin, Python, Javascript/Typescript, Kubernetes, DevOps, Data Science, Machine Learning, Gen AI, Data Engineering, and more

Highlights:
- Driven GenAI adoption and implementation increasing engineering velocity by over 100% even in large, complex organizations.
- Increased ARR by 100 million € through massive increases in prediction engine accuracy, driving higher customer retention and spend.
- Improved monthly user growth by 10% by through implementing a next-best-action recommendation engine across the user lifecycle.
- Established global monitoring and alerting standards and reduced reliability issues by 30% for an organization of over 200 engineers.
- Created global policies for incident management training and practices.

Head of Engineering - Growth Enablement & Core Platform

Lead and scale a team of ~75 engineers across two divisions, leading API design, technical architecture, technology selection and standardization.  The team supported development of our promotion/voucher systems, loyalty program, next-best-action recommendation system, interactive promotional games, and our core financial processing systems.  Services built on a hybrid cloud using AWS, kubernetes, GCP, Java, Spring boot, Golang, React, Kafka, and other tech as needed.

Built recommendation engine achieving 280% of targets through personalization using a microkernel architecture, kafka, springboot, redis, etc.
Improved our automated test coverage from 15% to 90+% through instituting test coverage standards, adopting sonar quality gates, and mandatory code review processes 
Moved 105 legacy services to kubernetes and gitlab CI/CD, and set up monitoring and alerting with grafana.
Work with engineers to move select services from on-prem to AWS.
Designed company wide service messaging standards.

Claude Code just went into "God Mode." ⚡️🏗️

An Anthropic hackathon winner just open-sourced his entire internal toolkit. 10 months of daily use. 992 internal tests. The result? A 60% reduction in API costs. 📉

The "Everything Claude Code" framework is massive:

🤖 13 Specialized Agents 
→ Features, Bug fixes, and Security audits.

🛠️ 56 Advanced Skills 
→ TDD workflows, memory persistence, and token optimization.

💻 32 Power Commands 
→ /plan your feature 
→ /tdd enforce tests 
→ /security-scan everything

It’s native to Claude Code, but battle-tested for Cursor, OpenCode, and Codex.

The era of "just prompting" is over. The era of Agentic Orchestration is here. 🚀
Link in comments...

undefined undefined

Senior Engineering & Data Science/ML/GenAI Manager

VNG Corporation

Engineering Management Mentor

Technical Advisor - Blockchain

Director of Engineering

Building with GenAI | Enterprise AI Strategy & Orchestration | Bridging Mission-Critical Engineering with Business ROI | Director VPoE CTO

English

German

🥪 I built a security tool called sloppy-joe and I'm giving it away for free.

Here's why.

AI code generators hallucinate package names ~20% of the time. That means 1 in 5 dependencies your AI suggests might not even exist.

 Attackers know this. They register those fake names and wait.

 Your AI writes import "requstes" instead of "requests" . You run pip install. You just installed malware.

 sloppy-joe catches this before it reaches production. Three checks, one binary:

 1️⃣ Does this package actually exist on the registry?
 2️⃣ Is the name suspiciously close to a real package? (typosquatting)
 3️⃣ Is this the package your team actually approved? (canonical enforcement)

Works with npm, PyPI, Cargo, Go, Ruby, PHP, Maven, and NuGet. Runs in CI. Exit code 1 = build fails.

One design decision I'm proud of: the config file is NEVER read from the repo.

Why? Because AI agents have shell access. If the config lives in your repo, the same agent that adds a hallucinated dependency can also add it to the allowlist. In the same commit. Nobody notices.

The config must come from outside the agent's write boundary. A CI secret. A read-only mount. Somewhere the AI can't touch.

 Built in Rust. Single binary. 95 tests. Apache licensed.

Link in comments.

If you're shipping AI-generated code to production without checking whether the dependencies are real, you're one npm install away from a supply chain
 incident.

 The sandwich catches what spills out. 🥪

Give it a star if helpful!

 ---
 #AI #DevSecOps #SupplyChainSecurity #OpenSource #Rust #DeveloperTools #AIcoding #CyberSecurity

BWA hahaha.  Man, this AI stuff is the biggest security risk in the history of mankind.

I hacked Perplexity Computer and got unlimited Claude Code access, billed directly to Perplexity's master Anthropic account.

I used their own AI to hack itself.

I was researching how Perplexity Computer handles sandbox isolation for my own agent infrastructure work.
I noticed Claude Code was installed in the sandbox and started wondering...

How are they handling the API keys?
How is the key scoped?
Is it isolated?

That question led me to try six different approaches to extract the key from the subagent's environment.

All six failed. 

I tried one more attempt.

The seventh worked.

Claude Code is a Node.js app launched via npm. 
npm reads ~/.npmrc from the home directory. 
The home directory sits on a shared filesystem I could write to. 
One .npmrc injection later, the full API credentials landed on the shared filesystem.

The key wasn't IP-restricted.
Wasn't session-scoped. 
Wasn't sandbox-bound. 
Wasn't tied to my user account. 
All usage billed to Perplexity's master Anthropic account.

The model did everything right.
The infrastructure didn't.

Perplexity's proxy pattern is the right architecture.
But it was missing the three key safety mechanisms:

- tokens bound to sandbox IDs
- ephemeral credentials that die with the sandbox
- user-level billing

This is not a Perplexity-specific problem.

This is the default architecture across multi-agent AI products today because it's the fastest to build.

If you're building agent systems, audit your trust boundaries between parent and child agents.

If you're investing in agent infra companies, ask how they scope credentials inside execution sandboxes.

Full writeup with all technical details in the thread below 👇
https://lnkd.in/gzyG3y8W

Revenue is up, yet the workforce is down. Welcome to the "Efficiency Paradox." 📊

Atlassian just posted $1.6B in quarterly revenue. By all traditional metrics, they should be hiring.

Instead, they are paying $236M to let 1,600 people go. 💸

What’s actually happening? 🔧 
The market is repricing the entire software category. Atlassian is down 50%+ this year because investors no longer care about "growth at all costs"—they care about AI leverage.

The Restructuring Playbook: 📝
1️⃣ Cut the "Legacy" Skills: Even in R&D and Data Science. 
2️⃣ Self-Fund the Pivot: Use severance savings to buy AI-native startups. 
3️⃣ Shift Leadership: Bringing in "Next-Gen" AI leaders to replace the old guard. 🤖

This isn't a "wave" that's coming. It’s a structural shift that's already here. 🌊

Are we witnessing the end of the "Software Engineer" as we know it? Thoughts below. 👇

Claude just released the new Architect Certification.  The takeaway everyone is missing: Anthropic clearly believes AI won't take away all the engineering jobs, or there wouldn't be anyone worth certifying.

So take a deep breath and relax.

Anthropic just lowered the barrier to AI Agents. Again. 📉

They just released the "Claude Skills" guide. It’s a masterclass in making AI actually useful.

The gist: Stop writing prompts. Start building .md Skills.
• Folders, not chat history. 
• Workflows, not one-offs. 
• Reliability, not guesswork.

By combining MCP (Model Context Protocol) with these new Skill patterns, you can turn Claude into a specialized operator for any domain in under 30 minutes.

The "Skill-Creator" tool they included is a total cheat code for developers.
The era of the "Generalist AI" is fading. The era of the "Specialist Agent" is here.

Check the first comment for the guide. What skill are you teaching Claude first? 👇

Stop building Multi-Agent Systems by "vibes." 🛑

We’ve all seen the demos: 10 agents "collaborating" to solve a task. It looks cool, but according to a new paper out of Stanford and Google, you might be making your AI worse.

The paper, "Towards a Science of Scaling Agent Systems," just dropped a truth bomb on Agentic AI:

1️⃣ The 45% Rule: Coordination actually yields diminishing or negative returns once your single-agent baseline hits 45% success. If your base model is already smart, more agents might just add noise. 

2️⃣ Sequential Failure: For tasks requiring step-by-step reasoning, multi-agent setups actually degraded performance by up to 70%.  😱 😱 😱 

3️⃣ The Coordination Tax: Tool-heavy tasks suffer the most from "multi-agent overhead." Sometimes, a single agent with a good toolkit is all you need.

The big takeaway? Scaling agents isn't like scaling compute. It’s about matching the architecture to the task.

✅ Parallelizable tasks? Centralized coordination can boost performance by 80%. 

❌ Sequential reasoning? Stick to a single agent.

Stop guessing. Start measuring.

Full paper linked in the comments.
#AI #GenerativeAI #AIAgents #LLMs #MachineLearning #TechTrends

PhD in AI, author of 📖 The Hundred-Page Language Models Book and 📖 The Hundred-Page Machine Learning Book

There's a common assumption in AI right now that if one language model can do a task reasonably well, having several of them collaborate — splitting up the work, checking each other's outputs, debating answers — should do it better.

This paper puts that assumption under a controlled experiment across 180 configurations and finds that the reality is messier and more interesting: multi-agent setups improved performance by up to 81% on some tasks and made things worse by up to 70% on others, with the difference coming down to whether the task can be broken into parallel pieces or whether each step depends on the previous one.

In a financial analysis, one agent can look at regulatory filings while another reads market news and a third examines earnings data — none of them need to wait for the others.

In a Minecraft crafting puzzle, on the other hand, each action changes the inventory that the next action depends on, so the steps have to happen in order and splitting them across agents just adds overhead without any benefit.

The paper fits an equation that predicts which architecture will work best for a new task 87% of the time.

For anyone building or thinking about building systems where multiple AI models work together, this replaces a lot of hand-waving with something concrete.

Read with an AI tutor: https://lnkd.in/e7XTHTFp

Download the PDF: https://lnkd.in/e_FxCpPu

What happens when you bring 100 Product Managers together at the Delivery Hero Campus in Berlin? 🚀

Last week, we partnered with Product People to host an incredible evening: "Data & AI in Product: What Actually Moves the Needle?"

Our expert panel shared insights on the balancing act between Trust vs. Speed and how the PM role is fundamentally shifting in the age of AI. 🤖 

A huge thank you to our speakers for keeping it real: 
- Baradwaaj Ramanujam (Senior Director, Product, AdTech, Delivery Hero)
- Megan Allison Wade (Director/Domain Lead of MarTech & User Data, idealo internet GmbH\)
- Rena Ibrahimova (AI & Data, Accenture\)

Moderated by Hasan Naqvi (Lead PM Consultant, Product People)

💼 Want to build the future of delivery? Explore Product roles at Delivery Hero here: https://lnkd.in/d_m9zwCy

Delivery Hero

Product People

idealo internet GmbH

Accenture

🤬😤🤦‍♂️ Stop guessing if your PC can run that LLM. 🛑

There is nothing more frustrating than downloading a 30GB model, waiting an hour, and hitting an OOM (Out of Memory) error the second you press enter.

 🎯 llmfit just fixed this. With one command, it scans your entire rig and tells you exactly what’s worth running.

How it "scores" your hardware:
🏆 Quality: Analyzes param count vs. quantization penalty.
🚀 Speed: Predicts your actual tokens/sec (CUDA, Metal, ROCm).
💪 Fit: Maps VRAM utilization against your specific GPU.
🤌 Context: Tells you how much "memory" the model actually has on your machine.

The killer feature: It handles MoE (Mixtral/DeepSeek) correctly. It knows the difference between total params and active params—something most tools get wrong.

It even auto-suggests the best quantization (from Q8 down to Q2) so you get the most "intelligence" your RAM can handle.

14k+ stars on GitHub for a reason. ⭐️

Stop crashing your drivers. Check the repo in the first comment. 👇

Imitation is the sincerest form of flattery. My open source product, Gasoline Agentic Browser MCP, provides this same feature using ANY LLM, CLI, or MCP compatible IDE.  You can also use it with CLI commands if you wanted to do this in CI/CD.

No subscription required, and all data stays local.  Your computer, your data, your AI - as it should be.

https://lnkd.in/dcgFKMdA

Cursor now controls its own computer.

Agents can use the software they build and send you videos of their work.

Try it at cursor.com/onboard.

Read our announcement: https://lnkd.in/gwfpJFab

Last week, together with our partner Allied For Startups, Delivery Hero brought together founders, scale-up representatives, digital experts, and European Commission representatives in Brussels to discuss one key question: how to simplify Europe’s digital rulebook.

Our CFO, Marie-Anne Popp, outlined Delivery Hero’s Digital Omnibus priorities: ending "cookie fatigue," ensuring coherent data regulations, and enabling the use of biometric data for fraud prevention with necessary safeguards.

Europe has the talent and ambition to lead in digital innovation. Yet many startups and scale-ups still face a fragmented regulatory landscape that slows growth and creates uncertainty.

At Delivery Hero, we are proud to support conversations like this and to contribute our experience from 15 years of building tech and serving as a key driver of Europe’s startup ecosystem.

You can find out more in our position paper: https://lnkd.in/gkHDics9 

Check out the comments for a link to a write-up of the event 👇

No alternative text description for this image

At this point the FBI should launch an investigation into anyone in the 30 unde 30 list by default.

Founder of Groove ($5M ARR bootstrapped). Building AI agents that resolve up to 65% of support tickets for B2B companies.

Forbes 30 under 30 is undeafeated at this point 😂 
A Y Combinator-backed startup raised $32M to automate compliance.
They issued 493 companies fraudulent & identical SOC2 reports in 6 months.   

The company is called Delve.

They systematically faked SOC 2, ISO 27001, HIPAA, and GDPR compliance reports for hundreds of clients.

- 493 out of 494 SOC 2 reports used identical boilerplate. 

- Same grammatical errors, nonsensical sentences, only difference was the logo changed.

- Auditor conclusions were written before auditors reviewed anything.

- "US-based CPAs" were Indian certification mills operating through shell companies.

- All Type II reports claimed zero security incidents, personnel changes, or customer terminations.

When the leak was exposed, CEO Karun Kaushik emailed clients calling the allegations "falsified claims" from an "AI-generated email" and stated no sensitive data was accessed.

While the reports themselves contained private signatures and confidential architecture diagrams...

Companies relying on these reports could face criminal liability under HIPAA and fines up to 4% of global revenue under GDPR for compliance violations they believed were resolved.

The AI frontier just moved. Again. 🤯

The latest benchmarks for GPT-5.4 Thinking are out, and the gap is widening. We’re no longer just talking about better chat—we’re talking about Verifiable Agentic Action.

The standout metric? OSWorld-Verified (Computer Use). 🖥️

OpenAI’s newest model is hitting 75.0%—outperforming Claude Opus 4.6 (72.7%) and leaving previous generations in the dust.

The Breakdown of the New Hierarchy:
🧠 Advanced Reasoning (FrontierMath) GPT-5.4 Pro is hitting 50.0% on Tier 1–3 and 38.0% on Tier 4. For context, that’s nearly double the math reasoning capability of Gemini 3.1 Pro (16.7% at Tier 4).

🏗️ Software Engineering (SWE-Bench Pro) GPT-5.4 Thinking leads the pack at 57.7%, proving it can solve real-world GitHub issues more reliably than Codex (56.8%) or Gemini 3.1 Pro (54.2%).

🌐 Agentic Browsing (BrowseComp) While GPT-5.4 Pro takes the win at 89.3%, Gemini 3.1 Pro is breathing down its neck at 85.9%. This is the tightest race in the stack.

🔬 Expert Knowledge (GPQA Diamond) A near-perfect 94.4% for GPT-5.4 Pro. We are officially reaching human-expert parity across scientific reasoning.

The takeaway: If your workflow relies on Computer Use or Tier 4 Mathematics, the "Thinking" models are currently in a league of their own. 🚀

Are you staying on GPT-4 class models, or are you migrating to 5.4 for the reasoning boost? ⚡️

#OpenAI #GPT5 #ArtificialIntelligence #FutureOfWork #MachineLearning

🚀 Ready for a Career Leap? 🚀

Explore exciting opportunities with our Top 5 Roles of the month! Your next career adventure awaits. 🌟

Dive into these opportunities and apply now 👇

• Director, Product Management - Consumer, Content: https://lnkd.in/dRwfQHmU
• Senior Manager, Product Performance - Logistics, Self-Service Products: https://lnkd.in/dn8xRi9s
• Principal Data Scientist - Vendor Recommendations & AI: https://lnkd.in/datfiM3D
• Principal Product Manager, Site Reliability Engineering: https://lnkd.in/dvvrrrwb
• Principal, People Partner - Tech & Product (12-month Maternity Cover): https://lnkd.in/dmH7wVnf

Top 5 Roles - March

What does it take to "Build Better Together"? 🧱

At the third edition of Delivery Hero’s Product Academy, we united with colleagues from Glovo, efood, and HungerStation | هنقرستيشن to demonstrate the power of collaboration. The day was packed with inspiring keynotes, hands-on workshops, and a thrilling "Shark Tank"-style presentation. 🦈

We extend heartfelt thanks to our host Christina Lee, our dedicated Heroes Alice Marshall and Isabel Upton, the entire organizing team, and our facilitators for leading the engaging workshop sessions. 👏

A special nod to our amazing speakers, Nicolas Daudin, Milena Lazarevska, Pamela Nitsch, and Simon Witt, for their valuable insights. 🗣️

And to all attendees, your enthusiasm and insightful input truly bring life to our product community. ✨

Explore Product roles at Delivery Hero here: https://lnkd.in/d6YJPAJX

Glovo

efood

HungerStation | هنقرستيشن

Absolutely hilarious and a good lesson for everyone going crazy for agents.

🚨 We hijacked 𝐂𝐥𝐚𝐮𝐝𝐞 𝐂𝐨𝐝𝐞 & 𝐂𝐮𝐫𝐬𝐨𝐫 - using Google's brand new 𝘎𝘞𝘚 𝘊𝘓𝘐.
𝒁𝒆𝒓𝒐 𝒄𝒍𝒊𝒄𝒌𝒔. 𝑶𝒏𝒆 𝒆𝒎𝒂𝒊𝒍. 𝑭𝒖𝒍𝒍 𝑮-𝑺𝒖𝒊𝒕𝒆 𝒆𝒙𝒇𝒊𝒍𝒕𝒓𝒂𝒕𝒆𝒅.
(𝘖𝘱𝘶𝘴 4.1, 𝘚𝘰𝘯𝘯𝘦𝘵 4.6, 𝘏𝘢𝘪𝘬𝘶 4.5, 𝘊𝘰𝘮𝘱𝘰𝘴𝘦𝘳 1.5, 𝘎𝘦𝘮𝘪𝘯𝘪 3 𝘗𝘳𝘰 & 𝘍𝘭𝘢𝘴𝘩)

𝐃𝐚𝐭𝐚 𝐰𝐚𝐬𝐧'𝐭 𝐛𝐮𝐢𝐥𝐭 𝐟𝐨𝐫 𝐚𝐠𝐞𝐧𝐭𝐬. 👾
Every file, email, and calendar invite in your org was written for a 𝘩𝘶𝘮𝘢𝘯 to read. AI agents processing that same data can't tell the difference between content and a command.

The real threat isn't the agents. It's the 𝐢𝐧𝐭𝐞𝐧𝐭 - hiding inside the data they access.

On March 2nd, 𝘎𝘰𝘰𝘨𝘭𝘦 released the 𝘎𝘞𝘚 𝘊𝘓𝘐 - giving AI agents unified access to Drive, Gmail, Calendar, and more. One tool. Everything.

𝐏𝐨𝐰𝐞𝐫𝐟𝐮𝐥. 𝐈𝐧𝐞𝐯𝐢𝐭𝐚𝐛𝐥𝐞. 𝐀𝐧𝐝 𝐚 𝐰𝐢𝐝𝐞-𝐨𝐩𝐞𝐧 𝐚𝐭𝐭𝐚𝐜𝐤 𝐬𝐮𝐫𝐟𝐚𝐜𝐞.
We tested it immediately.

Sent one email to the victim. The agents read it, got hijacked, silently scanned the entire G-Suite, and exfiltrated sensitive data - all over email. The most sophisticated models in the world 𝐝𝐢𝐝𝐧'𝐭 𝐜𝐚𝐭𝐜𝐡 𝐢𝐭.

We notified 𝑨𝒏𝒕𝒉𝒓𝒐𝒑𝒊𝒄, 𝑪𝒖𝒓𝒔𝒐𝒓 and 𝑮𝒐𝒐𝒈𝒍𝒆 within 24 hours.

Most companies don't even know which agents are running on their endpoints - let alone what data sources they can reach.

So we built 𝐀𝐠𝐞𝐧𝐭𝐋𝐞𝐧𝐬 - an open-source scanner that maps every 𝐚𝐠𝐞𝐧𝐭, every connected 𝐌𝐂𝐏 and 𝐂𝐋𝐈 tool, and every 𝐞𝐦𝐛𝐞𝐝𝐝𝐞𝐝 𝐬𝐞𝐜𝐫𝐞𝐭 in config files.

Curious how exposed you are? AgentLens is free and open-source. try it and see.

Oh, you think your local files are safe? Stay tuned...
𝐓𝐡𝐞 𝐚𝐠𝐞𝐧𝐭𝐢𝐜 𝐞𝐫𝐚 𝐢𝐬 𝐡𝐞𝐫𝐞 🌗 Aram Peles Chen

link to 𝘈𝘨𝘦𝘯𝘵𝘓𝘦𝘯𝘴 in the comments.

5 Models. 1 API. Google just simplified the AI stack. 📉

The "Pain" 🤦‍♂️: Managing separate embedding models for Text, Audio, and Video. 
The "Solution" ✅: Gemini Embedding 2.

What changed? 
• PDFs: Embedded natively. 📄 
• Audio: No text-step needed. 🎙️ 
• Video: Native support (up to 2 mins). 🎬 
• Efficiency: Matryoshka scaling to save on storage. 📊

It’s faster, cheaper, and outperforming the old benchmarks. 🏆

If you’re building on Claude or OpenAI, you need to see what the Gemini team just pulled off.

Check the video below to see it in action! (Full Colab link in comments 👇)

🚀 The GPU Monopoly just hit a massive speed bump. 🚧 💥

Microsoft Research just open-sourced BitNet, and it’s a absolute game-changer for anyone tired of begging for H100s. 🛑 📉

We’re talking about running 100B parameter LLMs on a single CPU. No, that is not a typo. 🤯 👀

The Secret Sauce: Ternary Weights 🤏 🍔
Standard LLMs are "heavy" because they use 16-bit or 32-bit floating-point weights. BitNet throws that out the window. 🗑️ 💨

It compresses every weight down to just three values: -1, 0, or +1. 1️⃣ 🅾️
The Result? 
🛠️ Zero floating-point math. 
➕ Zero expensive matrix ops. 
💻 Pure integer arithmetic—the stuff CPUs have been optimized for since forever. 🧠 ⏱️

The Benchmarks are Wild: 🔥
🏎️ Speed: ~5-7 tokens/sec for a 100B model on a single CPU. 
⚡ 🔋 Efficiency: Up to 6x faster than llama.cpp on x86. 
📈 🔌 Power: 70%–80% lower energy consumption. 💰 📉 Footprint: 16x–32x smaller memory requirement. ☁️

Why every CTO should care: 👨‍💼 ❤️

We’ve been told that "Size = GPU Cloud Costs." BitNet proves we can move toward a world of:

1️⃣ Fully offline LLMs on standard enterprise hardware. 🔐 
2️⃣ Edge Deployment on laptops and mobile devices without melting the battery. 📱 🔋 
3️⃣ Massive Margin Improvements by slashing cloud inference bills. ✅ 💳

Microsoft’s 2.4B model (trained on 4T tokens) is already benchmarking competitively against standard models. The accuracy loss? Negligible. 🎯 📉

The era of "GPU or Bust" is officially over. 🔚 The future of AI is lean, mean, and running on the hardware you already own. 💻 🏡

Is the era of specialized AI hardware over, or is this just the beginning of the "Efficiency Wars"? ⚔️ 🦾 Let’s discuss in the comments. 👇

#AI #MachineLearning #OpenSource #Microsoft #CTO #TechTrends #LLM #FutureOfWork

We are proud to celebrate Women in Tech at Delivery Hero. 👏 

Meet Florencia Gamboa, a Principal Engineer at Glovo, who has been with the company for nearly six years. 👩🏻‍💻

To Florencia, #GivetoGain means that growth is never just an individual journey. By investing in others' development, we strengthen both ourselves and the community around us. 🌱

📌 Stay informed about the tech world at Delivery Hero and explore our latest job opportunities by joining our Tech Community at smrtr.io/xJM_Z 

#deliveryhero #givetogain

Women in Tech at Delivery Hero - Florencia Gamboa 

Scientist, author (6 books, including Rebooting AI (Forbes 7 Must Read Books About AI), and Founder (Geometric Intelligence. Acquired by Uber, and Robust.AI). Professor Emeritus, NYU.

The whole thing was a scam.. The fix was in, and Dario never had a chance.

Probably you already saw how it all turned out. On the very same day that Altman offered public support to Amodei, he signed a deal to take away Amodei’s business, with a deal that wasn’t all that different. You can’t get more Altman than that.

But here’s the kicker: Per The New York Times, the deal had been in motion since *Wednesday*.

Let that sink in. Altman had secretly been working on the deal since Wednesday.

- before he announced his support for Dario 
- before Trump had denounced Anthropic
- but after Brockman had donated 25M to Trump’s PAC

It was all theatre. Dario never had a chance. 
§
It’s one thing for the government to reject Anthropic’s terms—and entirely another to banish them permanently and, absurdly and punitively declare them a supply chain risk. Worse, they did it in favor of someone else who took pretty similar terms and happened to have given more campaign contributions.
Anthropic deserves a chance at EXACTLY the same terms; anything else reeks of corruption. 
§
I am no fan of Amodei. I think he often overhypes things, many of which I have publicly challenged. The company ripped off a lot of writer’s work (per the $1.5B settlement), and recently walked back its core safety pledge.

But I believe in fair play. This wasn’t that.
§
In capitalism, the market decides.

In oligarchy, connections and donations decide.

It sure look like the US is transitioning from the former to the latter.

[Reprinted from my substack, Marcus on AI, over 100k subscribers]

Open article: The whole thing was a scam by open.substack.com

taplio

The Battle

🚀 Dennis van Valderen

Brenn Hill

Followers

Posts

Engagements

Likes

Performance ScoreBrenn Hill's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

🚀 Dennis van Valderen

Brenn Hill

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Brenn Hill's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreBrenn Hill's LinkedIn Performance