Nicolay Gerold

Not sure whether Codex or Devin are saving me time. The workflow is off. They are made to push to the remote, which completely breaks the workflow (working directory -> staging -> remote). So you push or make a pr, pull it down, check whether it's working or ...

make some small adjustments and push again (or rather just discard it). It makes sense that it is working locally, where the dev is working. I think it also creates a sense of detachment from the code. Usually when I pull something, it's the work of someone else....

@howaiisbuilt.fm is back. most AI conversations obsess over capabilities. This one focuses on constraints - the right ones that make AI actually useful rather than just impressive demos. open.spotify.com/episode/1MbZ...

EU mobilizes 200 billion Euros for AI. Unless we have a massive political change that money will just go to waste. For innovation to happen, we don't need money first, but deregulation. We cannot work on breakthrough technologies, when we are at constant fear of being sued.

We also need to celebrate the breakthroughs we make here in the EU. In the last few days politicians started to celebrate OpenAI moving to Munich as a breakthrough for the EU and technology in Germany.

EU mobilizes 200 billion Euros for AI. Unless we have a massive political change that money will just go to waste. For innovation to happen, we don't need money first, but deregulation. We cannot work on breakthrough technologies, when we are at constant fear of being sued.

- Safe harbor provisions for data experimentation within defined boundaries - R&D tax credits from day one - Reduced social security contributions for the first X employees - Fast-track banking setup and simplified financial reporting

EU mobilizes 200 billion Euros for AI. Unless we have a massive political change that money will just go to waste. For innovation to happen, we don't need money first, but deregulation. We cannot work on breakthrough technologies, when we are at constant fear of being sued.

Otherwise we set them up for failure from the get-go. We need easy ways to open up a new company. Otherwise we don't incentive people to fund, because it's just easier to take high-paying jobs from US startups or big-tech. We need tax exemptions for new startups for the first year.

EU mobilizes 200 billion Euros for AI. Unless we have a massive political change that money will just go to waste. For innovation to happen, we don't need money first, but deregulation. We cannot work on breakthrough technologies, when we are at constant fear of being sued.

We are creating fights on unequal footing. US startups can compete without the same regulations. Incumbents can compete with a bunch of lawyers figuring out ways around regulations.

EU mobilizes 200 billion Euros for AI. Unless we have a massive political change that money will just go to waste. For innovation to happen, we don't need money first, but deregulation. We cannot work on breakthrough technologies, when we are at constant fear of being sued.

The regulation EU startups are facing (GDPR, AI Act, DSA,...) are great in spirit, but when startups have to think regulation first, they are forced to move slower than the non-EU competitors.

RAG is dead. Long live RAG. LLMs suck at long context. This paper shows what I have seen in most deployments. With longer contexts, performance degrades.

So no more matching, but actual reasoning. Even top AI models struggle as texts get longer. - o3-mini degrades to 18.9 at 32K context window - Deepseek R1 to 20.7 at 32K - Most models even struggle past 4k For most tasks good retrieval + smaller contexts will improve your performance.

RAG is dead. Long live RAG. LLMs suck at long context. This paper shows what I have seen in most deployments. With longer contexts, performance degrades.

NoLima is a benchmark to test if AI models can truly understand long texts. - Insert one key fact (called a "needle") into a long text of irrelevant information - Ask the AI to find and use this fact to answer a question - The twist: The question and the key fact use different but related words

RAG is dead. Long live RAG. LLMs suck at long context. This paper shows what I have seen in most deployments. With longer contexts, performance degrades.

People say long-context is solved, because we perform well on "needle in the haystack". We place an answer somewhere in the context window and ask a question. When the model can answer correctly we can say, the model works at the given context length.

I am dreaming up an Excel backend. Excel as the frontend, Iceberg as the backend. But how do you handle formulas, pivot tables, ...? How do you malformatted sheets? You probably have to put restrictions on what you can do with Excel so it's valid to be stored.

Dropping some new episodes on @howaiisbuilt.fm . Links below.

New podcast with Alex Garcia on search in sqlite