Steffen Hoellinger

Many of the coding wrapper companies have recently changed their monetization to move almost exclusively to usage based. It’s a necessary step for these companies to be GM+ but it will also accelerate the economic incentive for the underlying Foundational Models to “move up the stack” as well. Otherwise, the FMs risk being disintermediated by agents/tool calls. Based on the recent valuations of the FMs, it’s unlikely that they can afford to let this happen. Next few quarters will be really interesting.

@martin_casado @levie Fully agree. Yet the early agent frameworks seem to be overloaded with complexity and the experience suffers from too many inefficient handovers.

.@levie made a great observation. Agent use is going counter to the simplistic AGI narrative of fewer, powerful agents with increasingly high level tasks. Rather, we're trending to more agents given narrowly scoped, well defined and narrow tasks. Generally by professionals.

@andreasklinger @pmoe Well deserved!

Yo. @pmoe just won the Investor of the Year away by the German startup association! That guy is too humble, doesn't even post about it – so i will brag for him. Philipp is one of the most active and best early stage investors in Europe and now leveling up Tiny aka TSIC (Tiny Supercomputer Investment Company, yes that's the full name) to become one of the leading VC early stage funds in Europe. No marketing, no PR releases or PR stunts, no swagger skitrips. He is the person that invests early in people and is also one of the first one's founders call if there is an issue. And by having some of the best numbers in the scene, he proves that his model works. Kudos mate!

💡 Real-time AI agents are only as intelligent as the data they process. I joined the "Life Is But A Stream" podcast with Joseph Morais (@thedatagiant) to talk about how real-time data streaming powers agentic AI and intelligent copilots. 📺YouTube: bit.ly/4jTSTGj 🎧Apple: bit.ly/43fUT4A 🎙️Spotify: open.spotify.com/episode/0Acfgv…

💡 Real-time AI agents are no longer just an idea...they’re here. Tune into the latest Life Is But A Stream episode with Joseph Morais (@thedatagiant) and Airy's Steffen Hoellinger as they dive into how real-time data streaming powers agentic AI and intelligent copilots. 📺 Watch it on YouTube: cnfl.io/459HVYp 🎧Listen Here: cnfl.io/42oI5JH

@martin_casado "There was never any more inception than there is now." (Walt Whitman, 1892)

My rough view on current models and AGI: - it's best measured economic value in a productive (not speculative) market - markets are adaptive - markets will shift to valuing new things - the universe is fractal, chaotic, non-linear and heavy tailed - there is an unlimited number of valuable new things to do (per @DavidDeutschOxf) - Current model architectures are a ways off from efficiently tackling new, heavy tailed spaces tl;dr expect humans to be intimately involved in high-value work for a very, very long time.

Full Talk: ververica.academy/courses/3d1634…

Summary of my recent #FlinkForward talk "Building #Copilots with #FlinkSQL, LLMs & Vector DBs": 🤖 AI Assistants & self-serve AI/BI for streaming & batch data 🧠 AI-powered schema intelligence 🔄 Natural Language to generate Flink SQL jobs ⚙️ Continuous monitoring for agentic AI

@martin_casado Just the beginning. Wait until your AI starts purchasing its own dev tools!

Ironically, AI seems to be increasing the dev tools TAM rather than decreasing it.

That's why you should ♥️ Logs.

𝐄𝐯𝐞𝐫𝐲 𝐒𝐲𝐬𝐭𝐞𝐦 𝐢𝐬 𝐚 𝐋𝐨𝐠 An idea on how to drastically 𝐫𝐞𝐝𝐮𝐜𝐞 𝐜𝐨𝐦𝐩𝐥𝐞𝐱𝐢𝐭𝐲 𝐚𝐧𝐝 𝐜𝐨𝐨𝐫𝐝𝐢𝐧𝐚𝐭𝐢𝐨𝐧 in distributed apps. A fun way to look at this is similar to the idea of Turning the Database Inside Out - like Turning the Microservice Inside Out. The core idea is: Applications coordinate with many systems (DBs, queues, locks, schedulers, ...) which are all logs underneath. But each log independently maintains its ground truth. We lack common order and conditions. By letting all those logs build on top of a common log, we can make a lot of distributed systems problems in apps and microservices virtually disappear. But how is this practically usable, given that our DBs and queues aren't built like this? How do we strike a sweet-spot balance between this model with its great resilience and consistency, and maintaining healthy decoupling and separation of concerns? We wrote about that design pattern, how it inspired our work on @restatedev, and how you can adopt this for your architecture. 👇

@euaccofficial Important one.

New proposal added to the euacc.com manifesto: 🇬🇧 Make English the primary language of the European Union English should become the primary language of the European Union because it is the only language that the majority of Europeans can understand or speak at least at a basic level. Unlike German or French, which are limited to specific regions and are not widely spoken outside their native countries, English has become the de facto lingua franca of Europe due to its use in education, business, and international communication. Clinging to languages like German or French for administrative purposes excludes a significant portion of the population and creates unnecessary barriers. Adopting English as the primary administrative language would promote efficiency, unity, and inclusivity across the European Union, reflecting the reality of how Europeans already communicate with one another. English should also become the official second language in every EU member state, making it mandatory to write and speak well in all schools from a young age.

Success is uncertain, but entertainment is guaranteed! ✨

@martin_casado Would love to join again!

Hey infra folks. We're standing up a new Discord server to discuss CS infra. If you want an invite DM me (reply and I'll follow). thanks!

@chrija @nathanbenaich Clear case for a multi-agent LLM system. Let's just have a virtual meeting where the notary AI reads out the financing documents via text-to-speech and all stakeholders' agents transcribe it back, process and re-negotiate if necessary. Best deals negotiating last minute changes!

When you've been talking about AI for ten years and your claim to fame is a tweet about the ridiculous notary procedure in Germany that went viral. 🤣 cc @nathanbenaich

@andreasklinger @nathanbenaich @skagarroum Great news! Still sad that Europe lacks private companies making transformative AI data center investments—yet.

Seven new high performance AI data center launching in Europe – funded by the EU Vibeshift? 🤞 ec.europa.eu/commission/pre… cc @nathanbenaich @skagarroum

@andreasklinger The magic word is “Bezugsurkunde” - instead of unproductively sitting around for hours (or even for days), rather have the notary read everything out to his/her own employees; then just sweep in for a 5 minute chat to give your blessing to the sacred ritual that had happened. 🚀

Young people have absolutely no respect for notary rituals anymore. They keep arguing that it’s pointless at company formation and only costs money and time. Can you believe this? This is an ancient ritual of our people and they dismiss it. Notary rituals are not about efficiency or business. They are about the elders of your community witnessing your new endeavors and acknowledging your existence. This is a ancient spiritual ritual involving a sacred cantation of the shareholder agreement in the blessed rite of incorporation. Your job is to sit. Be silent. And nod while you contemplate about your decisions that got you here and further. You do this for one hour until the cantation is done. Blessed is the omniregistry god and the SHA spirit. 🙏 How can young people be against this?!

The major areas of AI innovation automate white-collar work. Reviewing the BLS’ data on employment for white collar work, I aggregated the data to these categories. It’s striking that most of them already have a significant number of AI startups pursuing their ambitions to change workflows. Software engineers were the first to benefit with Copilot. Today, there’s a panoply of different kinds of AI software for developers, including test generation, code refactoring, code generation, & security analysis. Within education, The promise of computer adaptive testing and an individualized tutor and the style of Alexander the Great & Aristotle is possible., Automated document ingestion & reconciliation for closing books is changing accounting. AI powered financial due diligence rips through public filings, private placement memoranda, & compliance automation. Legaltech has surged automating demand letters for personal injury, writing briefs for attorneys, & chomping away paralegal work. We could keep iterating through the list. But if a founder wanted a list of jobs to automate, the BLS’ white collar jobs boards is a wonderful place to start. AI will change these workflows & capture a meaningful fraction of the labor spend. It’s a roadmap for the White Collar Revolution. tomtunguz.com/the-white-coll…

@sarahcat21 We give it stream processing as a powerful tool; works with weather data as well. 😉

If you just gave an LLM access to weather data (maybe injected it into context), would its performance improve and across what tasks?

@martin_casado Strongly agree. Increasing model scale also introduces greater ambiguity due to variations and inconsistencies in the training data. Ideally, proof of correctness should reside outside the model.

You can't fix hallucinations with model scale. Even reducing them by orders of magnitude still requires full error handling if correctness is required. No amount of GPUs or data will magically fix that. Like the Internet, we need to move up the stack for formal guarantees. There are eery parallels to the early Internet where the fiber folks where focused on building out capacity. And disconnected from the software folks how new that applications hadn't matured enough to take advantage of best effort. So, yes model progress may slow down because we're running out of data. But I suspect the real limitation is going to be at the app layer. It takes a long time for the industry to evolve its application stack.